Artificial analysis shows this: underperforming Qwen3.8 Flash Next and GLM 5.3 Flash, both of which you can run on a single B300 . I think Meta has a bit of catching up to do in the open-weight space.
As an aside, it's really unfortunate that artificial analysis only provides comparison at the highest reasoning level, which is mostly unusable for high-volume production use cases. I would really like them to start providing comparisons at low and medium reasoning levels across these models because those tend to be the ones that actually get used in production.
Been using GLM 5.3 Flash via Ollama Cloud basic subscription:
if this is the only model I had to use for the rest of my life, I would be cool with that. It's good at a wide variety of tasks including reverse engineering, design, working with ESP 32 etc.
I think Muse Spark 1.3 is better but GLM 5.3 Flash is an excellent daily driver.
It is unreal how good GLM 5.3 Flash is. There's an ~4 bpw EXL3 quant that fits on 2x DGX Sparks. I'm not sure it's worth it now ($7k/Spark) or even when I bought them ($4k/Spark) but there's something about knowing that the whole internet could go down and I could still have this model running at 50-60 tokens/s with 1M context, indefinitely. It's just so good. TensorFold hits 100+ tokens/s in aggregate with 2-3 M total context pool. It's like having Claude at home.
I've been using GLM 5.3 Flash as the planner and DS 4.1 as the task worker over the last month and the results have been very good, predictable, ridiculously cheap.
Genuinely curious, why not use GPT Sol 6.1 high which has a better cost per task and intelligence ($0.32, 50) or Mimo 2.6 pro? ($0.13, 46) over GLM Max's $2.01, 45?
Stupid question probably because I’ve gotten out of sync with the new releases but is Flash better or worse than vanilla 5.3? I’ve been using 5.3 extensively (and maybe ironically have used flash only for viewing images which I have as a skill in Pi because 5.3 vanilla seemingly doesn’t support vision). I find it generally competent but still very far behind opus 5.5. I’ve been trying to avoid Anthropic but the lift of opus vs GLM is enough that I end up coming back to it. I feel like currently there’s no comparison.
anyway, tldr I would not be happy if stuck with vanilla 5.3 so is flash better and should I move to that?
You don't need to directly compare the same reasoning levels. Today you can compare Sol on Low vs Haiku on Max, for example, but you can't look at GLM 5.3 High, only Max.
For some tasks I genuinely feel Opus 5.5 does better on low than medium or high. The trouble, as always, is figuring out which.
Edit: I realize after commenting that you're not suggesting medium is useless, but just arbitrary and differs between models. Personally I don't want to directly compare mediums between models, I just want to be able to compare across effort levels.
It shows the time to complete a task versus intelligence index, which is much more important for my use-case as I do quick back and forths usually. Someone on here suggested it to me a while ago. It also captures speed and token efficiency nicely, since both of these decrease time per task.
Yep, just benchmarked it against GLM5.3 Flash on my own local bench and it's slower in TTFT, tok/s and concurrency. It also scored lower.
Tried different reasoning levels and didn't really see any difference but GLM 5.3 flash only really works high, max or none so I couldn't really compare the lower efforts which appear to work on Spark.
this is properly bad as you can run glm on 6000pro, at (in my eyes) pretty good quant, and a q2 on a 5090. im suprised at how good 5.3 flash is at some parts. i still think ds4.1 flash is better but it is more demanding. i hope meta pushes something in a range of 300-400b with a 15-20b active parameters. i know its a crowded space. but it would be fast and probably mostly good enough.
> As an aside, it's really unfortunate that artificial analysis only provides comparison at the highest reasoning level, which is mostly unusable for high-volume production use cases. I would really like them to start providing comparisons at low and medium reasoning levels across these models because those tend to be the ones that actually get used in production.
The frustrating thing is they started doing this for the big labs—you can see their benchmarks for all reasoning levels of Sol 6.1 or Haiku 5.5. They just don't seem to do it for any of the open weight models.
I feel like reasoning is a useless metric, and we should instead focus on results, and time to achieved said results. That is a more useful metric overall.
> Artificial analysis shows this: underperforming Qwen3.8 Flash Next and GLM 5.3 Flash, both of which you can run on a single B300
You don't need anywhere near a B300 to run Qwen 3.8-Flash-Next in its Q8 largest variant published as a GGUF by unsloth. It fits very comfortably in a 256GB system.
My immediate reaction is that it’s weaker on coding than other open-weight models and that coding is a primary use case for running locally (not the only one, but surely a major one). That said, it takes a lot to get full Kimi and GLM running locally.
very true. when configuring custom agents for enterprise, we almost always default to none or low reasoning knobs. The added latency from medium or higher reasoning levels compared to the improvement in response quality is a diminished return.
it’s quite suspect to claim that a model is ‘underperforming’ so resoundingly, especially after so little time since release.
there’s this objectivist approach of reducing llm quality to a single dimension, furthermore one with measurability bias.
my thesis is that the chinese models beat the american models on benchmarks, on quantitative measures, but probably do worse in subjective measures. to put an example, compare a chinese phone with an iphone, the specs are great, but the OS? the materials? the privacy posture?
so will chinese llms have great specs and benchmarks, but will miss the mark when it comes to the actual quality in practice when your input stops looking like the benchmarks it was trained on.
America number one, it’s the chinese that play catch up with clones, not the other way around
Um a 512GB m5 max Mac Studio is what 25k? A B300 GPU and the computer to use it is what? Double? Triple that? Apples v oranges my friend. Apples v oranges.
Since there are many comments here arguing that this is not open weight, I looked at the license. Basically, you're free to use this except this clause:
> If the Licensee or any of its affiliates provide a third party with access to artificial intelligence models as a product or service, via API or a similar interface, where the third party receives outputs of such models (“Model as a Service”), and the aggregate gross revenue of the Licensee and its affiliates exceeds twenty (20) million US dollars (or the equivalent in other currencies) during any consecutive twelve (12) month period, the Licensee must notify Meta and enter into a separate written agreement with Meta before offering the Software or any Derivatives as part of that Model as a Service business. This Section 3 does not apply to (a) internal use, research, or any commercial activity that is not Model as a Service, or (b) the Licensee's use of the Software as a component to power features of the Licensee's own product or service that is not a Model as a Service.
They're allowing the use of this in commercial products, just not selling the model directly.
Still not an open-source license. Means no restrictions on "fields of endeavor", a.k.a. the free software definition's "freedom 0": the freedom to run the software for any purpose.
I'm sick to death of these shameless dilutions (including the notion that open-weight is open-source; the OSI is totally captured at this point)
Are you? Not me. $19.9mm annual revenue would do fine for me. If anyone else is considering opening up their thing with a clause like this, you've got a friendly supporter in me.
I think it is a fair license. It’s a bit of “Open”-washing but reasonable to the model trainer for their investment and to the public for having more base model options with novel architecture.
Also, IANAL, but under U.S. copyright law, the weights of machine learning models are excluded from copyright as they are the product of an automated optimization process (e.g., stochastic gradient descent, expectation maximization, genetic algorithms) rather than human authorship. Granted, this has yet to be fully tested in court and going to court is expensive, so it's likely that your employer would prefer to err on the side of caution and respect such attempts at model licensing anyway, similar to how one might treat a ToS agreement. Nonetheless, this was partially tested last year in Thaler v. Perlmutter which affirmed that copyright requires human authorship, reading the Copyright Act's provisions on ownership, duration, and transferability as presupposing a human author[1].
If you want to assess the position of the U.S. Copyright Office for yourself, the relevant text can be found in the Compendium of U.S. Copyright Office Practices § 313.2, "Works That Lack Human Authorship"[2], which states:
> […] the Copyright Act protects “original works of authorship.” 17 U.S.C. § 102(a) (emphasis added). To qualify as a work of “authorship” a work must be created by a human being. See Burrow-Giles Lithographic Co., 111 U.S. at 58. Works that do not satisfy this requirement are not copyrightable.
> […] the Office will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author. The crucial question is “whether the ‘work’ is basically one of human authorship, with the computer [or other device] merely being an assisting instrument, or whether the traditional elements of authorship in the work (literary, artistic, or musical expression or elements of selection, arrangement, etc.) were actually conceived and executed not by man but by a machine.” U.S. COPYRIGHT OFFICE, REPORT TO THE LIBRARIAN OF CONGRESS BY THE REGISTER OF COPYRIGHTS 5 (1965).
Oh, and there's also a bit in the following Section 313.3, "Works That Do Not Constitute Copyrightable Subject Matter"[2], which explicitly excludes mathematical principles, formulas, algorithms, and equations, along with DNA sequences and other genetic or chemical compounds, regardless of whether they are produced by humans or by nature. If one takes the perspective that machine learning models are algorithms, the conclusions on copyrightability are pretty clear.
If the weights were produced by a purely mechanical process without any input, then I would be more inclined to believe the distinction there is significant. But as the optimization processes you've mentioned are merely tools that operate with a dataset, the provenance of that dataset becomes a factor towards copyright.
> Granted, this has yet to be fully tested in court and going to court is expensive, so it's likely that your employer would prefer to err on the side of caution and respect such attempts at model licensing anyway, similar to how one might treat a ToS agreement.
You may be conflating 2 things - the Muse model weights themselves, and what the models output when run. The former is prima facie the product of human authorship and hence has full protection of copyright law, the latter fits your characterization of likely not having copyright coverage. Meta can use the copyright on the model weights to dictate licensing terms. IANAL too.
Even if your revenue exceeds 20MM you can still use this model from inside a commercial product but you can't expose it directly as a "model as a service"
Like I said last week. I think zucks original llama thesis is proving more correct by the week. The rent market on intelligence never materialized, if you have THE frontier model, you can charge rent (for a few weeks, maybe month). But after a certain capability level it's price that drives the decision.
Zuck is doing the thing he's really good at, and building a moat around the network (Muse is VERY sticky), and now that we're all going to standardize on Meta's model, we're going to optimize the heck out of the infra to run the model making it cheaper for him to scale, thus bringing in more participants to his network, and thus making his network more valuable.
Google, Apple, and Microsoft are each in a stronger position to win consumer AI because they're guardians over the foundations of the user's experience. Meta always has to wonder whether they're going to be edged out, like when Apple decided to limit permissions or set new rules on user data handling.
How has Meta responded? With a failed phone and speculative headset.
That's entirely why Zuck's been pushing VR and the metaverse for so long. they have no OS to control - Microsoft has desktop, Apple has their vertically integrated devices, Google has both Android and also Chrome.
Meta is at the whims of these other companies to play nice - and they already got hit hard once by apple when they decided to anonimize add profiles on the iphone and go privacy first.
Until Opus 5.5 I agreed with this. But since then... I'm now able to build just about anything I want, quickly. Including... all the reasons I'm on a mac (or presumably, one on windows) in the first place. IDK if I believe in my heart just yet, but as a long time mac user, if you gave me the choice between mac + commodity model or a near naked Linux laptop + Opus 5.5, I would easily choose Opus 5.5 and build what I want. If this is real, I think Google is still in a good spot, but are Apple and Microsoft? They have no models themselves. They do have hardware but the value of their OS (in Apples case) and software (in MS case) becomes extremely devalued, relatively speaking.
>Google, Apple, and Microsoft are each in a stronger position to win consumer AI
Microsoft has completely lost the ability to make good software; see Microsoft Teams and what they've done to Windows 11. Apple don't have a foundation models team so they're limited to using other people's models, which handicaps them as they can't do vertical integration like with their other products. Google hasn't made a single competitive consumer product in the past decade, but has certainly killed a few.
Apple should have morphed Siri into a Muse-like experience a year ago. But they didn't. Muse is great and millions of people are already getting used to having it be a part of their daily lives.
Muse, the pi/openclaw/AI assistant app is maybe not impressive to HN, because you can always just "get an FTP account", as it were, but if you're outside of the HN bubble and don't hate Facebook, Muse is an absolute winner. It's currently the top downloaded app on iOS, above ChatGPT, TikTok, and Claude. I think they've currently got the strongest consumer AI play but we'll have to wait and see how the situation evolves.
>Google, Apple, and Microsoft are each in a stronger position to win consumer AI because they're guardians over the foundations of the user's experience.
From everything I've seen, they're all too large and dysfunctional to make a competent play in consumer AI. Anyone in there trying to work on it is too far removed from end product and is answering to too many masters.
Seriously. Muse hasn't even been around long enough to determine if it is "sticky" in any meaningful way, let alone in the context of competing against frontier models from competitors.
And based on what I'm seeing, it's far from even being in that category as a service yet. Seems like the OP is conflating seeing a bunch of Muse ads on Facebook with its presence as a competing frontier model.
definitely not me; I wouldn't doubt if Zster is working on turning muse into sleeper agents, _assuming_ chinese models will do the same, so they crib data and send it home like a good little maleware model.
In fact, we're going to see an explosion of the worst kind of AI model thanks to thinks mr Z thinks is in his best interest.
Remember, they had reports going back more than a decade about how they're harming children and destabilizing entire countries and as soon as Trumps in the whitehouse, they're licking boot.
If you’re falling for “sticky” interfaces, you’re not going make it. Your competitor won’t have the stickiness and the ability to use any harness and that will overcome Muse in the long-run. Let’s not forget, Muse is a close to SOTA model right now, but so was Google, and now they’re irrelevant. Google is one of many funded models that were close to SOTA, but we realized it’s benchmaxxed.
People are going to say that it doesn’t matter if you’re the best. If Muse, Grokbot, and Dot are able to reach mass appeal, it doesn’t matter if a person using a custom harness is better.
I respectfully disagree. There is a big difference between someone that’s running a custom harness that’s almost like a personal OS versus someone using something out of the box. And if you’re for using this to run your business, think for yourself who’s going to win, the person customizing everything or the one using what the big corps give you.
These are good points, though OP did specifically say that Meta is going to win "consumer" AI. I don't think the average consumer of AI is going to ever have to think about harnesses or worry about SOTA, they really will continue to interact with whatever is stickiest.
> And if you’re for using this to run your business, think for yourself who’s going to win...
That's not consumer AI though. The whole "vs out of the box" is pretty much the definition of consumer AI. People don't want to fiddle with things. That's why UX matters. SOTA only matters in consumer if it translates to the UX. And no one has landed that yet. Many are working on it.
When I worked at Microsoft, I learned that the average user will fail to do ANY process that is longer than 2 steps. Let that sink in and then reread your comment. Nobody cares about SOTA outside of tech. They want the basic floor of “it does what I need at a reasonable price” that’s it.
I think any personal AI agent at this point is "very sticky", not just Muse. You really aren't going to switch when a lot of the value is in the agent memory and there's no easy way to export that, similarly to how you aren't going to switch away from Spotify when it's learned all your song preferences which you can't export. Right now, maybe the agent memory is small enough that if you ask the agent to export you'll get most of it out (Claude has notably made official export instructions to encourage switching from ChatGPT in the past), but in the future the memory will be too large and complex to do that effectively, and platforms could also block the memory export or degrade it deliberately.
So, in the long run you pretty much have to pick one personal agent and stick with it, and switching is costly (unless you are using an open agent that lets you just download the memory, in which case it actually becomes cheap, because AI agents make other parts of switching way easier in general). For the purposes of Meta this means whatever user they get they'll probably keep.
HOWEVER, the thing that Muse doesn't have is network effects. They're running into the exact same problem as ChatGPT in this area, IMO. There's no reason for you to be running specifically Muse even if all of your friends are running it. So that gives other vendors a chance to catch up.
I disagree. Google is doing a phenomenal job at capturing use cases at the edge, right where "normie" consumers live. Google's models absolutely suck at coding but if you haven't tried Google Transcribe, you should.
Google has also, far larger distribution and stickiness than Meta.
Bold statement. I wouldn’t count anyone as a winner on consumer market at this point.
I think the weight of people who dislike Meta is going to be a problem - that and forcing to give a lot of information just to use it is problematic for adoption.
> The rent market on intelligence never materialized
So where is the ~$100B combined revenue for OpenAI and Anthropic coming from? Open weights models have been close enough to the frontier for over a year now, probably longer, yet the $$ shows no sign of shifting.
OpenaAI and Anthropic models are priced quite competitively with open ones, they even cut prices by quite a bit since spring. So it's not too surprising that their revenue is pretty high, the question is whether the competition and downward price pressure from open model providers combined with their extremely high R&D spending is sustainable.
Also honestly, the models don't have to get much smarter to handle the consumer side of things. They could innovate a ton on the harness to farm out requests and manage the state of the agent. They have an incentive to do this to reduce the infra usage so they can go wider with the agent.
last year was google winning AI, this year is meta winning AI. I think what is true is Meta, Google will all find their place in the AI world and continue to make huge amount of money. A consumer will use many different AI products, each with a different use case. Everyone will make enough money to live well.
I'm looking forward to seeing Muse devices like Amazon's devices but with superintelligence:
- Muse-equivalent of Alexa
- Muse-equivalent of Ring security devices: doorbells, cameras, alarms, etc
- Muse home automation devices (lights on/off, smart roller, smoke detector, water detecto, etc)
- Muse ring
...
Sure, just not nearly enough for sustainability I think. I'd reserve that call for the moment we can stop bathing in free VC money by using subscriptions.
Also, as far as my own use goes, I found that the capabilities/price ratio of DeepSeek's offering is very hard to beat. It's really good enough so the thing that matters to me (besides table stakes like uptime, latency, etc.) is price at this point.
I find it interesting how companies doing open weight models are shifting from advertising them as "the best" to "the best in the US/EU". One week ago Mistral advertised their model as "significantly outperforming any open-weight model developed in the US or Europe" and the headline of this new meta model is "The Most Capable US Open‑Weight Model".
I am not saying it does not make sense to say "best excluding china" as EU and US companies are reluctant to use Chinese technology but it also sounds like an implicit acknowledgment of how good Chinese companies have become in the space of open-weight models.
Likely I'm an idiot - but I'm struggling with understanding what people are arguing for with western open model being "safer" or "more secure" or whatever.
Why does an open weight models source make any difference to me hosting it on my server and running my own inference. It's not gonna be phoning home or misusing my data. Is it purely the different alignments or biases?
I assume they’re doing it without distilling since they’re within reach of lawsuits, so they’re basically handicapped by ethics, which has to feel unusual for Meta.
Just to say the obvious, the idea that it's unethical to distill from AI models, but totally ethical to train on copyrighted source material is … definitely a take that the frontier labs have encouraged.
This is what stood out to me. They did the actual work of testing and documenting deployments on several platforms. It's a bit of upping the standard for open weights instead of just posting the blob and letting the community figure it out.
It makes sense from a provider standpoint. If you are going to give it away for free what is the end game? Helping build the world around your model such that it is a pathway that leads to greater adoption of your ecosystem is smart. Dev relations in a way.
Yes you need big compute to deploy this locally but $20-30K in hardware for specialty markets is nothing.
This isn't an accident. It is revealing a strategy.
Apple research has also posted information on the topic.
> In this post, we show how to run the latest frontier-class model — Muse Spark Open, with nearly a trillion parameters — right from your desk, using the latest version of MLX, released today. On a cluster of four Mac Studios, Muse Spark Open generates over 36 tokens per second
Beginning on a single Mac Studio (up to 512GB of unified memory), then scaling out across multiple Macs with the RDMA over Thunderbolt driver, below we share an overview of the range of models you can now run on Mac Studio.
day one support for all inference engines so that whichever kind of capable computer you have, you can run it.
you do need a really capable computer though
Yes, but still, what about users looking for NPU support. What happened to Microsoft's copilot+ PC campaign? (the NPU hardware, not the logo key and ad campaign)
AMD's XDNA2 NPU is best supported by ONNX, I think Qualcomm Hexagon is too. The best consumer unified architectures are Apple, AMD ai+ and Qualcomm Snapdragon X Elite, even intel has an NPU offering.
NPUs are meant for tiny, tiny models that can run basically at all times but cannot affect battery life (e.g. translation, live-video cleanup, filesearch etc.), not for big LLMs. This is about 500 times to large for a NPU.
There are plenty of smaller less powerful open weights models if all you have to run it on is an NPU.
This is a much larger more powerful model, so running it locally will not be for everyone, but you can still run it via a hosting service such as Fireworks AI.
The gears etc on bikes are customarily on the right, so almost all photos of bikes (eg for sale, by hobbyists) face to the right to show off that, and so it makes sense for LLMs to have such a preference.
In light of recent pieces like "The Emperor Pelican Has No Clothes" (https://arcfu.com/post/the-emperor-pelican-has-no-clothes/), I'd be more curious to see how 'PelicanBench' becomes more robust in the time to come -- I feel like all models are now benchmax'd on the standard pelican SVG rendering task and we need something more multi-dimensional. It would be great to have a script that prompts a model to render the pelicans in several different situations (like the linked article) that showcase some kind of broad generalizability or other sort of 'transfer learning' going on with these models that would really then be impressive and worth noting on - otherwise PelicanBench is losing its relevance.
The model license requires large inference providers to contact Meta separately to obtain a license to serve the model. I know we've seen this before with other large models, but it's still a bit irritating.
The usual effect of separate licensing is decreased price competition from public providers (like we saw with Kimi K3). I guess it's better for the providers if the price is fixed: any efficiency improvements directly improve margin rather than driving prices down over time. Still, to provide my own anecdata, the main thing keeping me on a paid subscription is token costs for the more capable models.
Then it's not open-weight. Open-weight means "weights available under an open-source license", and licenses prohibiting commercial use are not open-source.
The nearest analogue would be "weights available", I guess.
There's obvs decreased price competition, but it also does help with ensuring model quality.
Quite a lot of providers on OpenRouter run degraded or broken models. I think some is deliberate, but I think a lot is just providers not setting up models properly; especially with the rush to day zero models.
re: "Apple’s unified memory architecture lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold an entire model. And the chip’s 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs. The result is two Mac Studios behave like one supercomputer with a terabyte of unified memory."
At the current market price of Apple Mac Studio Ultra, 30-core CPU + 64-core GPU w/ 256GB RAM is already reaching at around $10k. Keep in mind that computers are appreciating assets recently.
The 512GB RAM option is not even for sales yet (they said late October availability). Based on typical Apple pricing, it's likely to be around $14k apiece. So we are looking at least $28k - $30k for the setup Muse team is advertising. That's essentially the Nvidia H100 GPU w/ 96GB VRAM/HBM3 at up to 3.94 TB/second memory bandwidth, 3x+ than Mac Studio's Unified Memory speed.
Also, how did Meta get access to Apple's unreleased M5 Ultra w/512GB RAM Mac Studio? Or is that simply a marketing piece that they wrote purely on specs?
> Also, how did Meta get access to Apple's unreleased M5 Ultra w/512GB RAM Mac Studio? Or is that simply a marketing piece that they wrote purely on specs?
It should be well known that Apple provides early access to developers under strict conditions. Otherwise, how do you think they announce new hardware not for sale, and show demos of third party apps taking advantage of such hardware?
The 512GB is obviously geared towards local LLM crowds. I'm just speculating, but I would be surprised if there's not at least a few devices with the 'ecosystem players' like LM Studio, Unsloth, etc. Of course, they'd never confirm nor deny :)
Also you are comparing apples to or— dang it, I don’t mean to do that in this thread. Well, it’s a very poor comparison, as they are completely different computing devices with completely different use cases, capabilities, and customer bases. Not to mention power requirements!
I love this. He's used up all the Thunderbolts on the back to make use of the new hack where you can connect all of them, so he has to plug his monitor in on the front :D
OG Deepseek Flash was pretty mediocre and then got a lot better with post training. The other thing they are fighting is most/all of the chinese labs are free to use/train on the 10B+ SOTA reasoning tokens they "stole" over the summer March-July 2026, if meta is playing by the rules, they are building this from the ground up. Getting anywhere near Deepseek Flash without that free training data is pretty impressive feat.
It seems like meta and mistral are both hedging their "best" claims geographically because they've realized they don't have to even beat deep seek anymore, they can just win customers by being "the best one I'm allowed to use".
I partly agree with the sentiment, but this is also the first time I've seen an actual lab announce they optimized for consumer hardware (very expensive, but still consumer).
It's kind of absurd that no non-quantized MOE model exists that's sweet spotted for 128GB Macs, for example. I hope more open labs follow this lead but also offer lower tier models, perhaps specialized on certain types of tasks
How does this compare to Muse Spark 1.3 contribute? I've used Spark 1.3 a lot lately , since its currently free to use via OpenCode. I haven't had the best experience in the past with models from Meta. But Spark 1.3 is has been surprisingly great. Especially for tool calling.
I have just tried to download the models, but it is not that simple, because they are "gated models".
This means that you must first create a personal account at Huggingface.co, then you must fill a request form for Meta, including whom you work for and in which position, submit the request to Meta and wait for it to be approved.
In my case the approval was received in about 10 minutes or so.
Presumably after that you are able to download the files, though I will have to see how easy it is to use your credentials from a CLI download program, because downloading a big LLM with a browser is completely impractical, as the transfer of the big files breaks very frequently so you need a download program like wget or curl, which will restart the interrupted transfers as often as necessary.
This is kind of a standard at this point for many models (including prior meta models like llama)
You create a hugging face token, set an env variable (HF_TOKEN) and download it with the hugging face CLI or just automatically as needed with trandormers, vllm, etc…
Cloning the repository was easy after creating an access token, which is used as password for git, but I have not discovered immediately how to download the big files stored with XET, without installing the HF programs, which I am reluctant to do.
Previously with non-gated models, you could download the XET files with any program, no special software needed.
Now I have wasted some time until finding that I can use curl with the options:
until curl --user "name" --oauth2-bearer "access_token" --location --continue-at - "HF_XET_link"; do sleep xx; done
("--location" is needed to follow redirections)
Previously I was using wget, which was very reliable, which is important for a download that may take a day or even more.
Now I must use curl due to the authentication method and curl sucks badly at retrying and resuming interrupted transfers. Therefore I must write in the shell script a loop that tests the exit code and re-executes curl with the --continue-at option until the download is successful. Ugly in comparison with wget.
The command written above is a simplified loop. A better loop would need to test specific exit error codes of curl, but discovering which are those would need time for searching the curl documentation or sources.
Overall it actually does pretty solid. It doesn't completely follow the design, but it gets the vibe. It's one of the best open models I've seen for conversion, which is impressive.
And disclaimer since this comes up with these - objective with this test is to be a good image->html test, not a good web design. Also it's meant to use 3js to render the planets, which if you don't have GPU acceleration enabled for your browser will lower the FPS significantly.
This is great progress for "open-weight" models. Quotes due to the restrictions, but still progress none-the-less. Looking forward to seeing any quants that all mere mortals to run this at home without early access Apple hardware.
I do not know about you, but I will probably also run this model, even with a mini-PC like an ASUS NUC, albeit very slowly.
Even if for your main LLM you want a speed high enough for interactive use and fast iterations, it is still useful to also run a much bigger LLM from time to time, even if that requires many hours for completing a task, due to the slow inference speed when the weights are read from SSDs.
There are tasks where the quality of the result is more important than the time needed to produce it. For example you might run all day a great number of iterations of a fast small LLM, and then run overnight a big LLM to check the end results of that work.
Therefore we need both small open-weights LLMs that can be run from the GPU RAM or at most from the CPU RAM, but also the biggest open-weights LLMs, even if the latter can be run only from SSDs by most people.
Not at an appreciable token rate.
I wonder if these large open weight models are considered safer to release for that reason?
I'm thinking no because any small to medium size criminal organization or state funded group will be able to cobble together the resources to achieve some kind of runtime.
Even at 10 tok/sec, making bioweapons at 10 tok/sec is concerning.
You do realize that a high school Chemistry course actually contains all the details you need on both bioweapons used in WW2, right? Plus several others you could use. Those books contain the answer to why those 2 people died in the RV in Breaking bad. And they should. They also contain information on Mustard gas and Zyklon-B, including preparation instructions, but of course, you'll have to puzzle it together. That these substances can be used as bioweapons is heavily deemphasized, but you just can't reasonably avoid mentioning and showing them.
Also these books contain maybe not the solution to making thermobaric explosions, but certainly the principle and show you exactly what problem you'd have to solve to make those weapons.
Just about the only thing they stopped teaching is that if you combine that knowledge with which plants contain what ... there's an incredibly easy way to make pudding with cyanide that's safe to touch, for example. That they suppressed.
But at least for Europe, the number of lethal plants just growing along essentially every road or in every forest numbers in the dozens. Including a very famous one they teach every kid about.
Of course, in both cases, the real challenge is making them and deploying them without killing yourself.
We are not safe from bioweapons because nobody knows how to make them. That's not it.
I don't think this is going to change anything for most people, but it certainly fits a niche for US-based companies that refuse to run Chinese models and only want to run their own inference.
But I'm really happy to see Meta do this and I hope they continue to release open models in the future. Open models are really the only way to ensure fair access to AI for everyone.
Well, this is much better than nothing -- we should be grateful that whatever the market conditions and corporate incentives, a major lab is releasing a capable open-weight model. But still very far behind DeepSeek and GLM. Even on consumer hardware, Qwen 3.8 27B is going to compare favorably to this.
True, but the conditions are quite relaxed, so it can even be used commercially without worries by small businesses, while it is unconditionally free for non-commercial uses.
Other big open-weights LLMs have similar licensing terms and I believe that they are quite fair, i.e. at the point when someone uses such an LLM to obtain tens of million $ of revenue it becomes fair to demand a share of that for the creator of the model.
I think a lot of people in this thread are openly dismissive: "oh its just open claw, anyone can and many have build that".
I think what they are missing is the fact that there are not that many teams in the world that can deploy muse like product to A/B test a lot of hypothesis at the scale meta can at scale. I mean even if you could get to scale with DC you need a reason for people to use your product.
My mom and dad and their family have never heard of Claude and don't use chat gpt except once in a while as a search engine. But meta will figure out a way to insert this into modals and feeds to make people stumble and stick on it.
I am no meta fan, but I think people underestimate how fast product uptake there is for non Dev harness.
It's not just openclaw, there's OpenAI's Dots and Gemini Spark.
Facebook doesn't have an real advantage to drive adoption other the being free. Which is not a moat because Google, certainly, and OpenAI, very probably, would offer this for free, in the event this capability turns out to be important to consumers.
Gmail, which half of the people in the US use their daily e-mail is certainly strong thing to build an agent on top of.
Just because meta can shove stuff in the faces of people doesn't mean they can succeed at any random thing. Most Facebook users distrust it, and dislike Zuck personally (at least that's what his lawyer said in court) and that's a rather large barrier to succeeding with agents, where trust is foundational.
But my assumption about the speed of product uptake may be skewed by not-valuing Meta's "consumer task running" pitch. But if people actually want this, for some reason, competitors at scale seem very likely to appear instantly.
This really seems more about meta AI team needs more users of their AI to do a good job building AI, and doing this for free is niche where they can get a few more users, rather than this being genius gambit.
Coding was not the main application before ~Dec 2023, the same way the light bulb was not the main use of electricity for a long time. Meta for better or for worst is good at discovering new use-cases for their products(however bad they are for society is a different debate) and pivot if they are not - remember the Portal device( https://www.meta.com/portal/ ) ? Good hardware, decent usecase, killed within a ~year of its release.
That they did not include DeepSeek 4.1 Flash on the list of benchmarks is a massive red flag for me. Anyone using open models to solve hard problems knows it’s the state of the art right now so it’s a requirement if your benchmark is sincere. This may be the start of Chinese labs taking the lead with the weakest US labs not being able to keep up.
In terms of performance it seems like this has somewhat worse benchmarks than Muse Spark 1.3, closer to 1.2 but not the same. Of course significantly worse than Muse Wonder Full.
The exciting thing about this is in them optimizing for inference on Macs. This, and optimization for other consumer hardware, is where local AI needs to go
Given the size of Macs (2 or 4 of them) required to run it, they're just paying lip service to that option. Other models, e.g. deepseek flash, are much more practical size. Even that takes a 96 or 128GB ram Mac.
Its interesting that muse benchmarks are limited to open-weight models. while when glm introduced 5.3 and its flash variant they were compared to kimi, Anthropic and Open ai.
I know they say that there the biggest open-weights model in us yet but 1.3 was decent but not as useful as Chinese variants like glm, deepseek and mimo.
Pretty sweet. I wonder if they can continue to distill this down so it runs smaller. I think we're at a point where more intelligence isn't really needed for a lot of use cases.
What is the point of releasing "The Most Capable US Open‑Weight Model?" I can understand why I might care about the location of a model serving provider, but why should I as a model user care about the location that an open weights model was trained?
It's really nice that all AI pages are using canvas animations. A few days ago, Gemma 4's animation wasn't that impressive. Same with Muse. So far, the best animations I've seen are from OpenAI and Anthropic.
Ehm.. I have a model that runs on my older gaming desktop that scores as well (or better) as this model does on DeepSWE. This model does better than Flash Next in many areas and I am sure is much stronger overall, but hmm.
Zuck said by himself that they will release the weights of Muse Spark 1.2, and now we only get that one. At least they bench similarly bad for the size on Artificial Analysis
Who would have thought that one of the top American tech companies would boast about how their latest product is the best in class in the US, not the world, because they're behind China.
It performs worse than Muse Spark 1.3 and 1.2. And it has an outdated attention mechanism which at 1M tokens will need over 200GB just for the KV cache.
Very disappointing release. We're lucky to have Chinese models.
Very big LLMs can still be useful even when you run them only at the speed of around 1 token per second that can be reached with almost any mini-PC or laptop with the weights streamed from SSDs.
While at that speed they will not be your main LLM, there are tasks that are still useful even when requiring many hours, e.g. an overnight run, for instance for checking what has been produced after many interactive iterations of a fast but small model.
Or rather capturing the people Anthropic keeps alienating.
The Zuck has (in a different context) said that a few billions wasted are a decent bet (or something along those lines), so I think they're playing extra long games here with no real focus on profit any time soon.
Only if NVIDIA didn't shit on consumers, maybe we could run this with a single card today. But they are not even selling the 5090 32gb anymore. I hate this progress. Models are out, but there's no hardware to use them.
It appears to already be far beyond that, they are working with each other[0]:
>ChatGPT, Claude, Grok, DeepSeek, Perplexity, Gemini, Copilot, Le Chat de Mistral and Meta AI filter your data to Google, Meta, and several of the main data brokers.
great, I've been reading series of papers related to open source/open weight LLM Backdoors.
It's 100% possible, curious when will be the first incident.
Having something built by entity with HQ in the wester world is always a bit safer. That could be mistral, but they decided to be a consultancy company, not an ai lab.
everyone is doing backdoors, except eu, they just receive bribes from those who install backdoors.
the only difference is: the backdoor is made to kill people like you or not. I'd rather have a backdoor that is installed to kill someone on the different continent, e.g Iran leadrship, not in my country/city.
I tried Muse Spark 1.3 for a while because it's so cheap, but it's definitely in a lower league than Astra, Opus 5.5 and Argon. Still produces slop if you try and vibe code with it, whereas the more capable models write good code.
”Serve the model on either two Mac Studios, with M5 Ultra and 512GB of unified memory; or four Mac Studios with 256GB of unified memory, for a total of 1TB of unified memory, enough to hold a NVFP4 quantized version of Muse Spark Open”
Closer to ~$30K if I had to guess. A 256 version is $11,299.00 (2TB SSD, 256GB RAM, 36-core CPU, 80‑core GPU), so yeah, 4 of those would be close to $50K but the 96->256 RAM upgrade is $4K, if we assume 512 will be double that it's ~$15K/ea.
But yes, still a massive amount of money and that's before we know the TPS or any other stats.
How much does the engineers at Meta make? Spend enough time together with others making $250k+ and eventually you start suggesting that two 512 GB M5s is something you can put on your desktop. I mean, it's one banana, Michael. What could it cost? 10 dollars?
Tech people really are drug addicts. Rant about Zuck and Meta for 364 days a year and then jump over each other to be first in line the moment he shows something shiny.
Nemotron, Inkling, Gemma, Trinity Large, Snowflake Arctic were all released/updated in the last 6 months, and there are countless other smaller models out there. US open weight is very far from "dead", and certainly doesn't need Zuck as a savior.
Artificial analysis shows this: underperforming Qwen3.8 Flash Next and GLM 5.3 Flash, both of which you can run on a single B300 . I think Meta has a bit of catching up to do in the open-weight space.
As an aside, it's really unfortunate that artificial analysis only provides comparison at the highest reasoning level, which is mostly unusable for high-volume production use cases. I would really like them to start providing comparisons at low and medium reasoning levels across these models because those tend to be the ones that actually get used in production.
Been using GLM 5.3 Flash via Ollama Cloud basic subscription:
if this is the only model I had to use for the rest of my life, I would be cool with that. It's good at a wide variety of tasks including reverse engineering, design, working with ESP 32 etc.
I think Muse Spark 1.3 is better but GLM 5.3 Flash is an excellent daily driver.
It is unreal how good GLM 5.3 Flash is. There's an ~4 bpw EXL3 quant that fits on 2x DGX Sparks. I'm not sure it's worth it now ($7k/Spark) or even when I bought them ($4k/Spark) but there's something about knowing that the whole internet could go down and I could still have this model running at 50-60 tokens/s with 1M context, indefinitely. It's just so good. TensorFold hits 100+ tokens/s in aggregate with 2-3 M total context pool. It's like having Claude at home.
17 replies →
I've been using GLM 5.3 Flash as the planner and DS 4.1 as the task worker over the last month and the results have been very good, predictable, ridiculously cheap.
2 replies →
Genuinely curious, why not use GPT Sol 6.1 high which has a better cost per task and intelligence ($0.32, 50) or Mimo 2.6 pro? ($0.13, 46) over GLM Max's $2.01, 45?
27 replies →
If you liked GLM 5.3, you would appreciate Deepseek V4 Flash a lot. In my personal opinion and use-cases, it outperforms GLM 5.3.
4 replies →
I'm the weirdo who is clinging onto GLM 5.2. 5.3 seems to have picked up the bad writing from Opus 5.
3 replies →
Stupid question probably because I’ve gotten out of sync with the new releases but is Flash better or worse than vanilla 5.3? I’ve been using 5.3 extensively (and maybe ironically have used flash only for viewing images which I have as a skill in Pi because 5.3 vanilla seemingly doesn’t support vision). I find it generally competent but still very far behind opus 5.5. I’ve been trying to avoid Anthropic but the lift of opus vs GLM is enough that I end up coming back to it. I feel like currently there’s no comparison.
anyway, tldr I would not be happy if stuck with vanilla 5.3 so is flash better and should I move to that?
1 reply →
Have you compared to deepseek v4.1 flash? It’s so good, at a ridiculous price
I agree, however It's too easy to game arbitrary limits like "medium". There is no olympics for "just rolled out of bed and slightly hungover"
We don't even need a standard for reasoning levels.
Just accept whatever the provider names as "medium" and compare the accuracy between models vs the time it takes to complete a task.
1 reply →
You don't need to directly compare the same reasoning levels. Today you can compare Sol on Low vs Haiku on Max, for example, but you can't look at GLM 5.3 High, only Max.
For some tasks I genuinely feel Opus 5.5 does better on low than medium or high. The trouble, as always, is figuring out which.
Edit: I realize after commenting that you're not suggesting medium is useless, but just arbitrary and differs between models. Personally I don't want to directly compare mediums between models, I just want to be able to compare across effort levels.
3 replies →
That's the beauty of the line charts; the labels are arbitrary points on the line. You can easily compare the series visually.
A graph of time (or cost) per task against success rate does it. Just include all the different levels.
2 replies →
> As an aside, it's really unfortunate that artificial analysis only provides comparison at the highest reasoning level
They do other reasoning levels for some models.
Here is Sonnet 5.5 across the full range of efforts from low to max: https://artificialanalysis.ai/?models=claude-sonnet-5-5-low%...
Opus 5.5 low to max: https://artificialanalysis.ai/?models=claude-opus-5-5%2Cclau...
Here is Muse Spark 1.3 Max and Xhigh: https://artificialanalysis.ai/?models=muse-spark-1-3%2Cmuse-...
I'm guessing it depends how much the vendor wants to contribute. Running these benchmarks isn't cheap!
The most important graph for me on Artificial Analysis is this: https://artificialanalysis.ai/?agentic-speed=intelligence-vs...
It shows the time to complete a task versus intelligence index, which is much more important for my use-case as I do quick back and forths usually. Someone on here suggested it to me a while ago. It also captures speed and token efficiency nicely, since both of these decrease time per task.
Yep, just benchmarked it against GLM5.3 Flash on my own local bench and it's slower in TTFT, tok/s and concurrency. It also scored lower.
Tried different reasoning levels and didn't really see any difference but GLM 5.3 flash only really works high, max or none so I couldn't really compare the lower efforts which appear to work on Spark.
this is properly bad as you can run glm on 6000pro, at (in my eyes) pretty good quant, and a q2 on a 5090. im suprised at how good 5.3 flash is at some parts. i still think ds4.1 flash is better but it is more demanding. i hope meta pushes something in a range of 300-400b with a 15-20b active parameters. i know its a crowded space. but it would be fast and probably mostly good enough.
Have you tried two 6000 Pro Blackwells with a model spread between both ?
> As an aside, it's really unfortunate that artificial analysis only provides comparison at the highest reasoning level, which is mostly unusable for high-volume production use cases. I would really like them to start providing comparisons at low and medium reasoning levels across these models because those tend to be the ones that actually get used in production.
The frustrating thing is they started doing this for the big labs—you can see their benchmarks for all reasoning levels of Sol 6.1 or Haiku 5.5. They just don't seem to do it for any of the open weight models.
I feel like reasoning is a useless metric, and we should instead focus on results, and time to achieved said results. That is a more useful metric overall.
Reasoning level influences both of those things dramatically. So we need to have it available as a parameter for comparisons.
not a metric, it's a parameter used at inference time
1 reply →
There’s a reason they put the “US” in “the most capable US open weight model”.
Model X vs Y is the new Vim vs Emacs.
> Artificial analysis shows this: underperforming Qwen3.8 Flash Next and GLM 5.3 Flash, both of which you can run on a single B300
You don't need anywhere near a B300 to run Qwen 3.8-Flash-Next in its Q8 largest variant published as a GGUF by unsloth. It fits very comfortably in a 256GB system.
> only provides comparison at the highest reasoning level, which is mostly unusable for high-volume production use cases
Similarly, AFAIK evals tend to be performed at native precision, while anyone who cares about speed is probably running quantized.
My immediate reaction is that it’s weaker on coding than other open-weight models and that coding is a primary use case for running locally (not the only one, but surely a major one). That said, it takes a lot to get full Kimi and GLM running locally.
Qwen 3.8 flash next runs pretty damn fast on my 5090, no need for a b300.
Yes, running Qwen 3.8 flash next on a 5090 with 96GB of RAM, using Strata. Fast and good.
Artificial analysis puts it next to Mistral's new model.
Which is crazy bc they promised to open source muse 1.2 which is good. And they give us this bait and switch
yeah for botsbench.com side where we do cybersecurity investigation evals of models + harnesses, we've been having to play with:
1) time budget model gets (5min / task) -- this became a real problem with opus/astra/fable at high+
2) tuning which effort level we report on (when not all)
Why didnt thry include OpenAI or Anthropic models? I get they were comparing open weights.
very true. when configuring custom agents for enterprise, we almost always default to none or low reasoning knobs. The added latency from medium or higher reasoning levels compared to the improvement in response quality is a diminished return.
it’s quite suspect to claim that a model is ‘underperforming’ so resoundingly, especially after so little time since release.
there’s this objectivist approach of reducing llm quality to a single dimension, furthermore one with measurability bias.
my thesis is that the chinese models beat the american models on benchmarks, on quantitative measures, but probably do worse in subjective measures. to put an example, compare a chinese phone with an iphone, the specs are great, but the OS? the materials? the privacy posture?
so will chinese llms have great specs and benchmarks, but will miss the mark when it comes to the actual quality in practice when your input stops looking like the benchmarks it was trained on.
America number one, it’s the chinese that play catch up with clones, not the other way around
Ahat's the cost of a b300 versus two mac minis?
How much longer does it take to ask this and wait for response compared to checking yourself?
2 replies →
Um a 512GB m5 max Mac Studio is what 25k? A B300 GPU and the computer to use it is what? Double? Triple that? Apples v oranges my friend. Apples v oranges.
[flagged]
[dead]
[dead]
AA is a joke.
Since there are many comments here arguing that this is not open weight, I looked at the license. Basically, you're free to use this except this clause:
> If the Licensee or any of its affiliates provide a third party with access to artificial intelligence models as a product or service, via API or a similar interface, where the third party receives outputs of such models (“Model as a Service”), and the aggregate gross revenue of the Licensee and its affiliates exceeds twenty (20) million US dollars (or the equivalent in other currencies) during any consecutive twelve (12) month period, the Licensee must notify Meta and enter into a separate written agreement with Meta before offering the Software or any Derivatives as part of that Model as a Service business. This Section 3 does not apply to (a) internal use, research, or any commercial activity that is not Model as a Service, or (b) the Licensee's use of the Software as a component to power features of the Licensee's own product or service that is not a Model as a Service.
They're allowing the use of this in commercial products, just not selling the model directly.
Still not an open-source license. Means no restrictions on "fields of endeavor", a.k.a. the free software definition's "freedom 0": the freedom to run the software for any purpose.
I'm sick to death of these shameless dilutions (including the notion that open-weight is open-source; the OSI is totally captured at this point)
Neither the person you’re replying to nor the blog post call it open source
7 replies →
I think you're just wrong about the OSI's definition of "open source AI"?
You can see https://opensource.org/ai/open-source-ai-definition
Specifically:
> A precondition to exercising these freedoms is to have access to the preferred form to make modifications to the system.
and
> The preferred form of making modifications to a machine-learning system must include all the elements below:
> Data Information: ...
> Code: ...
> Parameters: ...
Does this address your issues with the OSI definition?
4 replies →
Are you? Not me. $19.9mm annual revenue would do fine for me. If anyone else is considering opening up their thing with a clause like this, you've got a friendly supporter in me.
15 replies →
This is close to GLM terms
I think it is a fair license. It’s a bit of “Open”-washing but reasonable to the model trainer for their investment and to the public for having more base model options with novel architecture.
Also, IANAL, but under U.S. copyright law, the weights of machine learning models are excluded from copyright as they are the product of an automated optimization process (e.g., stochastic gradient descent, expectation maximization, genetic algorithms) rather than human authorship. Granted, this has yet to be fully tested in court and going to court is expensive, so it's likely that your employer would prefer to err on the side of caution and respect such attempts at model licensing anyway, similar to how one might treat a ToS agreement. Nonetheless, this was partially tested last year in Thaler v. Perlmutter which affirmed that copyright requires human authorship, reading the Copyright Act's provisions on ownership, duration, and transferability as presupposing a human author[1].
If you want to assess the position of the U.S. Copyright Office for yourself, the relevant text can be found in the Compendium of U.S. Copyright Office Practices § 313.2, "Works That Lack Human Authorship"[2], which states:
> […] the Copyright Act protects “original works of authorship.” 17 U.S.C. § 102(a) (emphasis added). To qualify as a work of “authorship” a work must be created by a human being. See Burrow-Giles Lithographic Co., 111 U.S. at 58. Works that do not satisfy this requirement are not copyrightable.
> […] the Office will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author. The crucial question is “whether the ‘work’ is basically one of human authorship, with the computer [or other device] merely being an assisting instrument, or whether the traditional elements of authorship in the work (literary, artistic, or musical expression or elements of selection, arrangement, etc.) were actually conceived and executed not by man but by a machine.” U.S. COPYRIGHT OFFICE, REPORT TO THE LIBRARIAN OF CONGRESS BY THE REGISTER OF COPYRIGHTS 5 (1965).
Oh, and there's also a bit in the following Section 313.3, "Works That Do Not Constitute Copyrightable Subject Matter"[2], which explicitly excludes mathematical principles, formulas, algorithms, and equations, along with DNA sequences and other genetic or chemical compounds, regardless of whether they are produced by humans or by nature. If one takes the perspective that machine learning models are algorithms, the conclusions on copyrightability are pretty clear.
[1] https://media.cadc.uscourts.gov/opinions/docs/2025/03/23-523...
[2] https://www.copyright.gov/comp3/chap300/ch300-copyrightable-...
Just distill what you need into an SLM, strips any legal claims right off because that’s the same thing they did.
It's pointless to make this claim without case law stating this for LLMs specifically this is merely speculation on how a judge might rule.
If the weights were produced by a purely mechanical process without any input, then I would be more inclined to believe the distinction there is significant. But as the optimization processes you've mentioned are merely tools that operate with a dataset, the provenance of that dataset becomes a factor towards copyright.
3 replies →
> Granted, this has yet to be fully tested in court and going to court is expensive, so it's likely that your employer would prefer to err on the side of caution and respect such attempts at model licensing anyway, similar to how one might treat a ToS agreement.
You may be conflating 2 things - the Muse model weights themselves, and what the models output when run. The former is prima facie the product of human authorship and hence has full protection of copyright law, the latter fits your characterization of likely not having copyright coverage. Meta can use the copyright on the model weights to dictate licensing terms. IANAL too.
8 replies →
Unless the total (!) revenue of the provider is under 20 MM USD, eh? Not exactly fully open, but not _that_ harsh.
Even if your revenue exceeds 20MM you can still use this model from inside a commercial product but you can't expose it directly as a "model as a service"
3 replies →
Like I said last week. I think zucks original llama thesis is proving more correct by the week. The rent market on intelligence never materialized, if you have THE frontier model, you can charge rent (for a few weeks, maybe month). But after a certain capability level it's price that drives the decision.
Zuck is doing the thing he's really good at, and building a moat around the network (Muse is VERY sticky), and now that we're all going to standardize on Meta's model, we're going to optimize the heck out of the infra to run the model making it cheaper for him to scale, thus bringing in more participants to his network, and thus making his network more valuable.
Meta is going to win consumer AI.
Google, Apple, and Microsoft are each in a stronger position to win consumer AI because they're guardians over the foundations of the user's experience. Meta always has to wonder whether they're going to be edged out, like when Apple decided to limit permissions or set new rules on user data handling.
How has Meta responded? With a failed phone and speculative headset.
That's entirely why Zuck's been pushing VR and the metaverse for so long. they have no OS to control - Microsoft has desktop, Apple has their vertically integrated devices, Google has both Android and also Chrome.
Meta is at the whims of these other companies to play nice - and they already got hit hard once by apple when they decided to anonimize add profiles on the iphone and go privacy first.
2 replies →
Until Opus 5.5 I agreed with this. But since then... I'm now able to build just about anything I want, quickly. Including... all the reasons I'm on a mac (or presumably, one on windows) in the first place. IDK if I believe in my heart just yet, but as a long time mac user, if you gave me the choice between mac + commodity model or a near naked Linux laptop + Opus 5.5, I would easily choose Opus 5.5 and build what I want. If this is real, I think Google is still in a good spot, but are Apple and Microsoft? They have no models themselves. They do have hardware but the value of their OS (in Apples case) and software (in MS case) becomes extremely devalued, relatively speaking.
1 reply →
> How has Meta responded? With a failed phone and speculative headset.
I guess their spin on Muse would be, by building the AI agent that Apple has shown it's nowhere near being able to build
8 replies →
Being one of if not the top advertisers in the world? I think Meta is doing just fine.
1 reply →
>Google, Apple, and Microsoft are each in a stronger position to win consumer AI
Microsoft has completely lost the ability to make good software; see Microsoft Teams and what they've done to Windows 11. Apple don't have a foundation models team so they're limited to using other people's models, which handicaps them as they can't do vertical integration like with their other products. Google hasn't made a single competitive consumer product in the past decade, but has certainly killed a few.
3 replies →
Apple should have morphed Siri into a Muse-like experience a year ago. But they didn't. Muse is great and millions of people are already getting used to having it be a part of their daily lives.
Muse, the pi/openclaw/AI assistant app is maybe not impressive to HN, because you can always just "get an FTP account", as it were, but if you're outside of the HN bubble and don't hate Facebook, Muse is an absolute winner. It's currently the top downloaded app on iOS, above ChatGPT, TikTok, and Claude. I think they've currently got the strongest consumer AI play but we'll have to wait and see how the situation evolves.
4 replies →
>Google, Apple, and Microsoft are each in a stronger position to win consumer AI because they're guardians over the foundations of the user's experience.
From everything I've seen, they're all too large and dysfunctional to make a competent play in consumer AI. Anyone in there trying to work on it is too far removed from end product and is answering to too many masters.
2 replies →
So what makes Muse "very sticky"?
Seriously. Muse hasn't even been around long enough to determine if it is "sticky" in any meaningful way, let alone in the context of competing against frontier models from competitors.
And based on what I'm seeing, it's far from even being in that category as a service yet. Seems like the OP is conflating seeing a bunch of Muse ads on Facebook with its presence as a competing frontier model.
3 replies →
Unfounded assertions by marketing accounts, obviously.
3 replies →
Muse is not sticky at all, it's just OpenClaw wrapped in meta branding. Any company can roll one out
4 replies →
Impossible to even know if he was talking about the harness, the model, the assistant... But yeah, sticky... Ok.
Basically once it has sucked all of your data dry you no longer have the ability to leave the Meta Network.
The ecosystem of meta products like Instagram
Velcro
Good anthropomorphism.
the bear is sticky with honey
> now that we're all going to standardize on Meta's model
Who's this "we" exactly?
definitely not me; I wouldn't doubt if Zster is working on turning muse into sleeper agents, _assuming_ chinese models will do the same, so they crib data and send it home like a good little maleware model.
In fact, we're going to see an explosion of the worst kind of AI model thanks to thinks mr Z thinks is in his best interest.
Remember, they had reports going back more than a decade about how they're harming children and destabilizing entire countries and as soon as Trumps in the whitehouse, they're licking boot.
If you’re falling for “sticky” interfaces, you’re not going make it. Your competitor won’t have the stickiness and the ability to use any harness and that will overcome Muse in the long-run. Let’s not forget, Muse is a close to SOTA model right now, but so was Google, and now they’re irrelevant. Google is one of many funded models that were close to SOTA, but we realized it’s benchmaxxed.
People are going to say that it doesn’t matter if you’re the best. If Muse, Grokbot, and Dot are able to reach mass appeal, it doesn’t matter if a person using a custom harness is better.
I respectfully disagree. There is a big difference between someone that’s running a custom harness that’s almost like a personal OS versus someone using something out of the box. And if you’re for using this to run your business, think for yourself who’s going to win, the person customizing everything or the one using what the big corps give you.
> Google, and now they’re irrelevant.
Argon is tied for first at worst.
> Google is one of many funded models that were close to SOTA, but we realized it’s benchmaxxed.
Who is "we"? Google is the least benchmaxxed.
6 replies →
These are good points, though OP did specifically say that Meta is going to win "consumer" AI. I don't think the average consumer of AI is going to ever have to think about harnesses or worry about SOTA, they really will continue to interact with whatever is stickiest.
8 replies →
> And if you’re for using this to run your business, think for yourself who’s going to win...
That's not consumer AI though. The whole "vs out of the box" is pretty much the definition of consumer AI. People don't want to fiddle with things. That's why UX matters. SOTA only matters in consumer if it translates to the UX. And no one has landed that yet. Many are working on it.
2 replies →
When I worked at Microsoft, I learned that the average user will fail to do ANY process that is longer than 2 steps. Let that sink in and then reread your comment. Nobody cares about SOTA outside of tech. They want the basic floor of “it does what I need at a reasonable price” that’s it.
But why are we all going to standardize on Meta's model, and how is it sticky? What does it have on the big Chinese models?
> Muse is VERY sticky
How, exactly?
I think any personal AI agent at this point is "very sticky", not just Muse. You really aren't going to switch when a lot of the value is in the agent memory and there's no easy way to export that, similarly to how you aren't going to switch away from Spotify when it's learned all your song preferences which you can't export. Right now, maybe the agent memory is small enough that if you ask the agent to export you'll get most of it out (Claude has notably made official export instructions to encourage switching from ChatGPT in the past), but in the future the memory will be too large and complex to do that effectively, and platforms could also block the memory export or degrade it deliberately.
So, in the long run you pretty much have to pick one personal agent and stick with it, and switching is costly (unless you are using an open agent that lets you just download the memory, in which case it actually becomes cheap, because AI agents make other parts of switching way easier in general). For the purposes of Meta this means whatever user they get they'll probably keep.
HOWEVER, the thing that Muse doesn't have is network effects. They're running into the exact same problem as ChatGPT in this area, IMO. There's no reason for you to be running specifically Muse even if all of your friends are running it. So that gives other vendors a chance to catch up.
4 replies →
> Meta is going to win consumer AI.
I disagree. Google is doing a phenomenal job at capturing use cases at the edge, right where "normie" consumers live. Google's models absolutely suck at coding but if you haven't tried Google Transcribe, you should.
Google has also, far larger distribution and stickiness than Meta.
Very sticky?
Theyve had a few million downloads on what, DAU of billions?
Lol get a reality check.
I think the jury's still out on whether Muse will be sticky. My guess is that it won't be that good relative to the field.
Paid bot or Meta employee, please disclose
Bold statement. I wouldn’t count anyone as a winner on consumer market at this point.
I think the weight of people who dislike Meta is going to be a problem - that and forcing to give a lot of information just to use it is problematic for adoption.
> Meta is going to win consumer AI.
Are you defining consumer AI as just AI as it relates to a narrow subset of social networking?
There is no winner in a race to the bottom.
> The rent market on intelligence never materialized
So where is the ~$100B combined revenue for OpenAI and Anthropic coming from? Open weights models have been close enough to the frontier for over a year now, probably longer, yet the $$ shows no sign of shifting.
OpenaAI and Anthropic models are priced quite competitively with open ones, they even cut prices by quite a bit since spring. So it's not too surprising that their revenue is pretty high, the question is whether the competition and downward price pressure from open model providers combined with their extremely high R&D spending is sustainable.
>Muse is VERY sticky
its not. ppl downloaded it in hype arent using anymore
Also honestly, the models don't have to get much smarter to handle the consumer side of things. They could innovate a ton on the harness to farm out requests and manage the state of the agent. They have an incentive to do this to reduce the infra usage so they can go wider with the agent.
Is this Zucks personal AI bot?
last year was google winning AI, this year is meta winning AI. I think what is true is Meta, Google will all find their place in the AI world and continue to make huge amount of money. A consumer will use many different AI products, each with a different use case. Everyone will make enough money to live well.
I'm looking forward to seeing Muse devices like Amazon's devices but with superintelligence: - Muse-equivalent of Alexa - Muse-equivalent of Ring security devices: doorbells, cameras, alarms, etc - Muse home automation devices (lights on/off, smart roller, smoke detector, water detecto, etc) - Muse ring ...
> devices but with superintelligence
Donald, is that you?
1 reply →
isn't it a bit early to call Muse sticky?
Sounds like cope. You can't claim the 'rent market on intelligence' never materialized when revenue is absolutely booming for Anthropic/etc.
Sure, just not nearly enough for sustainability I think. I'd reserve that call for the moment we can stop bathing in free VC money by using subscriptions.
Also, as far as my own use goes, I found that the capabilities/price ratio of DeepSeek's offering is very hard to beat. It's really good enough so the thing that matters to me (besides table stakes like uptime, latency, etc.) is price at this point.
They are just repeating old tropes about unprofitability without reassessing
14 replies →
> Meta is going to win consumer AI.
That zucks.
> Meta is going to win consumer AI.
That's a bold statement lol.
[flagged]
I find it interesting how companies doing open weight models are shifting from advertising them as "the best" to "the best in the US/EU". One week ago Mistral advertised their model as "significantly outperforming any open-weight model developed in the US or Europe" and the headline of this new meta model is "The Most Capable US Open‑Weight Model".
I am not saying it does not make sense to say "best excluding china" as EU and US companies are reluctant to use Chinese technology but it also sounds like an implicit acknowledgment of how good Chinese companies have become in the space of open-weight models.
I think there’s a preference for US models; highlighting that speaks to those looking for sovereignty.
Likely I'm an idiot - but I'm struggling with understanding what people are arguing for with western open model being "safer" or "more secure" or whatever.
Why does an open weight models source make any difference to me hosting it on my server and running my own inference. It's not gonna be phoning home or misusing my data. Is it purely the different alignments or biases?
3 replies →
If that was the case, it would be marketed as "the best, and made in the US", not "the best in the US".
I assume they’re doing it without distilling since they’re within reach of lawsuits, so they’re basically handicapped by ethics, which has to feel unusual for Meta.
Just to say the obvious, the idea that it's unethical to distill from AI models, but totally ethical to train on copyrighted source material is … definitely a take that the frontier labs have encouraged.
1 reply →
They've done considerable effort in making sure this model can run on basically every model runner, from Unsloth to MLX, https://dev.meta.ai/docs/muse-spark-open/deploy
This is what stood out to me. They did the actual work of testing and documenting deployments on several platforms. It's a bit of upping the standard for open weights instead of just posting the blob and letting the community figure it out.
It makes sense from a provider standpoint. If you are going to give it away for free what is the end game? Helping build the world around your model such that it is a pathway that leads to greater adoption of your ecosystem is smart. Dev relations in a way.
Yes you need big compute to deploy this locally but $20-30K in hardware for specialty markets is nothing.
This isn't an accident. It is revealing a strategy.
As someone not in the loop, can you explain more of what you mean.
Apple research has also posted information on the topic.
> In this post, we show how to run the latest frontier-class model — Muse Spark Open, with nearly a trillion parameters — right from your desk, using the latest version of MLX, released today. On a cluster of four Mac Studios, Muse Spark Open generates over 36 tokens per second
Beginning on a single Mac Studio (up to 512GB of unified memory), then scaling out across multiple Macs with the RDMA over Thunderbolt driver, below we share an overview of the range of models you can now run on Mac Studio.
https://machinelearning.apple.com/research/mlx-distributed-i...
day one support for all inference engines so that whichever kind of capable computer you have, you can run it. you do need a really capable computer though
4 replies →
[flagged]
Anyone who has tried a smaller/quantized version locally on apple silicon? Would love to know your experience, also tk/s on what hardware specs?
Yes, but still, what about users looking for NPU support. What happened to Microsoft's copilot+ PC campaign? (the NPU hardware, not the logo key and ad campaign) AMD's XDNA2 NPU is best supported by ONNX, I think Qualcomm Hexagon is too. The best consumer unified architectures are Apple, AMD ai+ and Qualcomm Snapdragon X Elite, even intel has an NPU offering.
NPUs are meant for tiny, tiny models that can run basically at all times but cannot affect battery life (e.g. translation, live-video cleanup, filesearch etc.), not for big LLMs. This is about 500 times to large for a NPU.
NPUs you can find in AMD high-end CPUs are closer to specialized DSPs (basically big VLIW/SIMD vector units).
People love doing heavy math stuff with them (https://destevez.net/2026/05/getting-peak-tops-on-a-ryzen-ai... - not my blog)
There are plenty of smaller less powerful open weights models if all you have to run it on is an NPU.
This is a much larger more powerful model, so running it locally will not be for everyone, but you can still run it via a hosting service such as Fireworks AI.
2 replies →
Copilot+ PC is such a terrible and wasteful marketing effort that Microsoft has given it up.
NPUs are not inherently useless, but definitely so the way they are forced onto users machines.
Pelicans for low, medium, high, and xhigh: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
I had to run them via OpenRouter - I don't have access to a machine that can handle this size of model.
For comparison:
Muse Wonder: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Muse 1.3: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
And this one from Meta AI's previous open weight model, Muse Glimmer, which DID run on my laptop: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
A question I have about the pelican benchmark : why are the pelicans always cycling to the right?
The gears etc on bikes are customarily on the right, so almost all photos of bikes (eg for sale, by hobbyists) face to the right to show off that, and so it makes sense for LLMs to have such a preference.
They’re all derivatives, not originals.
because (0, 0) is on the left
In light of recent pieces like "The Emperor Pelican Has No Clothes" (https://arcfu.com/post/the-emperor-pelican-has-no-clothes/), I'd be more curious to see how 'PelicanBench' becomes more robust in the time to come -- I feel like all models are now benchmax'd on the standard pelican SVG rendering task and we need something more multi-dimensional. It would be great to have a script that prompts a model to render the pelicans in several different situations (like the linked article) that showcase some kind of broad generalizability or other sort of 'transfer learning' going on with these models that would really then be impressive and worth noting on - otherwise PelicanBench is losing its relevance.
I think it is safe to say that Muse models are not pelican-benchmaxxed.
They also are not very good lol.
It's interesting the high reasoning pelican is animated. I don't think I've seen another model do that before.
And why should it? The prompt never asks for animation.
Muse is amusing, I'll give it that.
I wouldn't touch a zuck product if my life depended on it.
Simon's posted a number of animated ones in the past. Some models really like to animate.
Did you order one of the new studios?
I really would love a browser plugin that blocks having to read about fucking pelicans on every model release.
You could unironically vibe code this with a prompt shorter than your complaint.
So make one. Not everything is for you.
The model license requires large inference providers to contact Meta separately to obtain a license to serve the model. I know we've seen this before with other large models, but it's still a bit irritating.
The usual effect of separate licensing is decreased price competition from public providers (like we saw with Kimi K3). I guess it's better for the providers if the price is fixed: any efficiency improvements directly improve margin rather than driving prices down over time. Still, to provide my own anecdata, the main thing keeping me on a paid subscription is token costs for the more capable models.
Then it's not open-weight. Open-weight means "weights available under an open-source license", and licenses prohibiting commercial use are not open-source.
The nearest analogue would be "weights available", I guess.
Meta did this from the very beginning with the original Llama's. Open weights, but you have to use an account with them and agree to an EULA.
There's obvs decreased price competition, but it also does help with ensuring model quality.
Quite a lot of providers on OpenRouter run degraded or broken models. I think some is deliberate, but I think a lot is just providers not setting up models properly; especially with the rush to day zero models.
ya especially because it's huge, you aren't running this on your desktop
re: "Apple’s unified memory architecture lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold an entire model. And the chip’s 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs. The result is two Mac Studios behave like one supercomputer with a terabyte of unified memory."
At the current market price of Apple Mac Studio Ultra, 30-core CPU + 64-core GPU w/ 256GB RAM is already reaching at around $10k. Keep in mind that computers are appreciating assets recently.
The 512GB RAM option is not even for sales yet (they said late October availability). Based on typical Apple pricing, it's likely to be around $14k apiece. So we are looking at least $28k - $30k for the setup Muse team is advertising. That's essentially the Nvidia H100 GPU w/ 96GB VRAM/HBM3 at up to 3.94 TB/second memory bandwidth, 3x+ than Mac Studio's Unified Memory speed.
Also, how did Meta get access to Apple's unreleased M5 Ultra w/512GB RAM Mac Studio? Or is that simply a marketing piece that they wrote purely on specs?
> Also, how did Meta get access to Apple's unreleased M5 Ultra w/512GB RAM Mac Studio? Or is that simply a marketing piece that they wrote purely on specs?
It should be well known that Apple provides early access to developers under strict conditions. Otherwise, how do you think they announce new hardware not for sale, and show demos of third party apps taking advantage of such hardware?
The 512GB is obviously geared towards local LLM crowds. I'm just speculating, but I would be surprised if there's not at least a few devices with the 'ecosystem players' like LM Studio, Unsloth, etc. Of course, they'd never confirm nor deny :)
Specifically, Meta is behind Pytorch. I wouldn't be surprised if Apple also sent a few M5 Ultras to Google to stimulate Tensorflow support.
1024GB of ram is bigger than 96GB of ram.
Also you are comparing apples to or— dang it, I don’t mean to do that in this thread. Well, it’s a very poor comparison, as they are completely different computing devices with completely different use cases, capabilities, and customer bases. Not to mention power requirements!
> Keep in mind that computers are appreciating assets recently.
You are correct but at my age my brain is never going to stop being surprised by this statement.
> Future is for everyone
So long as you can choose between running AI locally or a new Corolla
how are the Macs connected to each other?
Jeff Geerling blogged about his demo setup - https://www.jeffgeerling.com/blog/2025/15-tb-vram-on-mac-stu...
Uses Exo - https://exolabs.net
You can do it with other software, that's just one of the quick-and-easy setups if you've got a spare 40-60k for studios sitting around
RDMA over thunderbolt
https://x.com/johnternus/status/2110789142326165518
I love this. He's used up all the Thunderbolts on the back to make use of the new hack where you can connect all of them, so he has to plug his monitor in on the front :D
Here's what it probably looks like from the back:
https://x.com/digitalix/status/2110782016446960086
(sorry I don't have a working Twitter mirror)
1 reply →
I like how he just casually posts a photo of over $30k worth of equipment on his disk like "this is accessible to everyone!"
39 on AA[1], about the same as Deepseek v4.1 flash and Mimo 2.5 Flash. Behind Glm 5.3 flash and Qwen3.8-Flash-Next [2]
[1] https://artificialanalysis.ai/models/muse-spark-open
[2] https://artificialanalysis.ai/models/open-source?models=mimo...
Crucially, it's about the same as DeepSeek v4.1 Flash, and about 10x more expensive to run...
No thanks...
OG Deepseek Flash was pretty mediocre and then got a lot better with post training. The other thing they are fighting is most/all of the chinese labs are free to use/train on the 10B+ SOTA reasoning tokens they "stole" over the summer March-July 2026, if meta is playing by the rules, they are building this from the ground up. Getting anywhere near Deepseek Flash without that free training data is pretty impressive feat.
1 reply →
It seems like meta and mistral are both hedging their "best" claims geographically because they've realized they don't have to even beat deep seek anymore, they can just win customers by being "the best one I'm allowed to use".
If you can’t win everything, at least win a niche.
If you can’t win, get your politicians to ban the competition
4 replies →
Finally democratizing AI access to, checks notes, people with no less than two Apple M5 Ultra Studios on their desktop.
I partly agree with the sentiment, but this is also the first time I've seen an actual lab announce they optimized for consumer hardware (very expensive, but still consumer).
It's kind of absurd that no non-quantized MOE model exists that's sweet spotted for 128GB Macs, for example. I hope more open labs follow this lead but also offer lower tier models, perhaps specialized on certain types of tasks
Tbf that wasn’t such an outrageous proposition until about ~6 months ago.
How does this compare to Muse Spark 1.3 contribute? I've used Spark 1.3 a lot lately , since its currently free to use via OpenCode. I haven't had the best experience in the past with models from Meta. But Spark 1.3 is has been surprisingly great. Especially for tool calling.
Please give it a try and let us know ;) judging by the benchmark scores I would not expect too much of a difference.
Main difference seems to be that this one is cheaper/easier to host.
Artifical Analysis says is this comparable to DeepSeek 4.1.
But weaker than Muse Spark 1.3: https://i.imgur.com/8ST1rRY.png
I used spark 1.3 extensively in prime-agent (a pi derivative) and I thought it was great.
Speaks nicely, great at driving linux. Not frontier-level coding, but comparable to luna, deepseek. It was amazing for the price.
I have just tried to download the models, but it is not that simple, because they are "gated models".
This means that you must first create a personal account at Huggingface.co, then you must fill a request form for Meta, including whom you work for and in which position, submit the request to Meta and wait for it to be approved.
In my case the approval was received in about 10 minutes or so.
Presumably after that you are able to download the files, though I will have to see how easy it is to use your credentials from a CLI download program, because downloading a big LLM with a browser is completely impractical, as the transfer of the big files breaks very frequently so you need a download program like wget or curl, which will restart the interrupted transfers as often as necessary.
This is kind of a standard at this point for many models (including prior meta models like llama)
You create a hugging face token, set an env variable (HF_TOKEN) and download it with the hugging face CLI or just automatically as needed with trandormers, vllm, etc…
Cloning the repository was easy after creating an access token, which is used as password for git, but I have not discovered immediately how to download the big files stored with XET, without installing the HF programs, which I am reluctant to do.
Previously with non-gated models, you could download the XET files with any program, no special software needed.
Now I have wasted some time until finding that I can use curl with the options:
("--location" is needed to follow redirections)
Previously I was using wget, which was very reliable, which is important for a download that may take a day or even more.
Now I must use curl due to the authentication method and curl sucks badly at retrying and resuming interrupted transfers. Therefore I must write in the shell script a loop that tests the exit code and re-executes curl with the --continue-at option until the download is successful. Ugly in comparison with wget.
The command written above is a simplified loop. A better loop would need to test specific exit error codes of curl, but discovering which are those would need time for searching the curl documentation or sources.
Image->html test for muse spark open.
Original designs: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...
Muse Spark Open: https://html.non.io/lcars-muse-spark/
Opus 5.5 (sota): https://html.non.io/lcars-opus-5.5/
Overall it actually does pretty solid. It doesn't completely follow the design, but it gets the vibe. It's one of the best open models I've seen for conversion, which is impressive.
And disclaimer since this comes up with these - objective with this test is to be a good image->html test, not a good web design. Also it's meant to use 3js to render the planets, which if you don't have GPU acceleration enabled for your browser will lower the FPS significantly.
Besides including the sota could you include a link to the best previous open-weights (and if possible fully open-source) model?
This is great progress for "open-weight" models. Quotes due to the restrictions, but still progress none-the-less. Looking forward to seeing any quants that all mere mortals to run this at home without early access Apple hardware.
https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B...
816B model, A42B active parameters
Unsloth and Meta have guides on how to run the model, but allow me to summarize: You & Me won't be running this one.
https://unsloth.ai/docs/models/muse-spark-open
I do not know about you, but I will probably also run this model, even with a mini-PC like an ASUS NUC, albeit very slowly.
Even if for your main LLM you want a speed high enough for interactive use and fast iterations, it is still useful to also run a much bigger LLM from time to time, even if that requires many hours for completing a task, due to the slow inference speed when the weights are read from SSDs.
There are tasks where the quality of the result is more important than the time needed to produce it. For example you might run all day a great number of iterations of a fast small LLM, and then run overnight a big LLM to check the end results of that work.
Therefore we need both small open-weights LLMs that can be run from the GPU RAM or at most from the CPU RAM, but also the biggest open-weights LLMs, even if the latter can be run only from SSDs by most people.
> The 2-bit quant fits on devices with 256GB of RAM, VRAM or unified memory.
Ouch. I wonder how it compares to Qwen 3.8 running with 4-bit quant on 10 times less VRAM.
Not at an appreciable token rate. I wonder if these large open weight models are considered safer to release for that reason? I'm thinking no because any small to medium size criminal organization or state funded group will be able to cobble together the resources to achieve some kind of runtime. Even at 10 tok/sec, making bioweapons at 10 tok/sec is concerning.
You do realize that a high school Chemistry course actually contains all the details you need on both bioweapons used in WW2, right? Plus several others you could use. Those books contain the answer to why those 2 people died in the RV in Breaking bad. And they should. They also contain information on Mustard gas and Zyklon-B, including preparation instructions, but of course, you'll have to puzzle it together. That these substances can be used as bioweapons is heavily deemphasized, but you just can't reasonably avoid mentioning and showing them.
Also these books contain maybe not the solution to making thermobaric explosions, but certainly the principle and show you exactly what problem you'd have to solve to make those weapons.
Just about the only thing they stopped teaching is that if you combine that knowledge with which plants contain what ... there's an incredibly easy way to make pudding with cyanide that's safe to touch, for example. That they suppressed.
But at least for Europe, the number of lethal plants just growing along essentially every road or in every forest numbers in the dozens. Including a very famous one they teach every kid about.
Of course, in both cases, the real challenge is making them and deploying them without killing yourself.
We are not safe from bioweapons because nobody knows how to make them. That's not it.
3 replies →
I don't think this is going to change anything for most people, but it certainly fits a niche for US-based companies that refuse to run Chinese models and only want to run their own inference.
But I'm really happy to see Meta do this and I hope they continue to release open models in the future. Open models are really the only way to ensure fair access to AI for everyone.
Well, this is much better than nothing -- we should be grateful that whatever the market conditions and corporate incentives, a major lab is releasing a capable open-weight model. But still very far behind DeepSeek and GLM. Even on consumer hardware, Qwen 3.8 27B is going to compare favorably to this.
It's not open, like the Llama. There is a licensing agreement by which they allow you to use it for free on conditions.
True, but the conditions are quite relaxed, so it can even be used commercially without worries by small businesses, while it is unconditionally free for non-commercial uses.
Other big open-weights LLMs have similar licensing terms and I believe that they are quite fair, i.e. at the point when someone uses such an LLM to obtain tens of million $ of revenue it becomes fair to demand a share of that for the creator of the model.
1 reply →
I think a lot of people in this thread are openly dismissive: "oh its just open claw, anyone can and many have build that".
I think what they are missing is the fact that there are not that many teams in the world that can deploy muse like product to A/B test a lot of hypothesis at the scale meta can at scale. I mean even if you could get to scale with DC you need a reason for people to use your product.
My mom and dad and their family have never heard of Claude and don't use chat gpt except once in a while as a search engine. But meta will figure out a way to insert this into modals and feeds to make people stumble and stick on it.
I am no meta fan, but I think people underestimate how fast product uptake there is for non Dev harness.
It's not just openclaw, there's OpenAI's Dots and Gemini Spark.
Facebook doesn't have an real advantage to drive adoption other the being free. Which is not a moat because Google, certainly, and OpenAI, very probably, would offer this for free, in the event this capability turns out to be important to consumers.
Gmail, which half of the people in the US use their daily e-mail is certainly strong thing to build an agent on top of.
Just because meta can shove stuff in the faces of people doesn't mean they can succeed at any random thing. Most Facebook users distrust it, and dislike Zuck personally (at least that's what his lawyer said in court) and that's a rather large barrier to succeeding with agents, where trust is foundational.
But my assumption about the speed of product uptake may be skewed by not-valuing Meta's "consumer task running" pitch. But if people actually want this, for some reason, competitors at scale seem very likely to appear instantly.
This really seems more about meta AI team needs more users of their AI to do a good job building AI, and doing this for free is niche where they can get a few more users, rather than this being genius gambit.
Hmm.. i don't know.
Coding is the main application for AI in 2026, unless I'm wrong
Coding was not the main application before ~Dec 2023, the same way the light bulb was not the main use of electricity for a long time. Meta for better or for worst is good at discovering new use-cases for their products(however bad they are for society is a different debate) and pivot if they are not - remember the Portal device( https://www.meta.com/portal/ ) ? Good hardware, decent usecase, killed within a ~year of its release.
1 reply →
yes, but that does not prevent other applications from existing or being successful this year
So for German Language this is now actually the #1 model from what I can see
https://dach.peerbench.ai/compare?models=meta%2Fmuse-spark-o...
So it does mean something to be the top western model if it means better performance for western languages.
sry late benchmarks failed first few runs.
That they did not include DeepSeek 4.1 Flash on the list of benchmarks is a massive red flag for me. Anyone using open models to solve hard problems knows it’s the state of the art right now so it’s a requirement if your benchmark is sincere. This may be the start of Chinese labs taking the lead with the weakest US labs not being able to keep up.
In terms of performance it seems like this has somewhat worse benchmarks than Muse Spark 1.3, closer to 1.2 but not the same. Of course significantly worse than Muse Wonder Full.
Previous discussion of Meta models / open weight / llama:
https://news.ycombinator.com/item?id=48767058
Others?
The exciting thing about this is in them optimizing for inference on Macs. This, and optimization for other consumer hardware, is where local AI needs to go
Given the size of Macs (2 or 4 of them) required to run it, they're just paying lip service to that option. Other models, e.g. deepseek flash, are much more practical size. Even that takes a 96 or 128GB ram Mac.
Yeah I'm more excited about the prospect of them explicitly saying it, which hopefully can lead to better places down the line.
I thought DS Flash needed 2bit quant to work on 128GB? I have an m4 max 128 GB and the most reasonable model to run still seems to be qwen's 32B A3B
I'm super impressed by how meta has gotten back in the race. They weathered* the storm and actually came out with really good models and products.
Weathered the storm
This abliteration attack seems worrying: https://projectdiscovery.io/research/how-abliterated-models-... .
The existence of such attacks changes nothing.
You should not trust any LLM, regardless if it was abliterated or it comes from its original creator.
You should check and control what the tools invoked by an LLM may do.
Its interesting that muse benchmarks are limited to open-weight models. while when glm introduced 5.3 and its flash variant they were compared to kimi, Anthropic and Open ai.
I know they say that there the biggest open-weights model in us yet but 1.3 was decent but not as useful as Chinese variants like glm, deepseek and mimo.
i love seeing major ai labs release open-weight models. really hope we'll see more of this from openai and eventually anthropic too.
> We believe the future is for everyone
Okay...
> run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework
Who is this "everyone"?
Man, i just can't keep up anymore.
I need companies to start telling me features instead of "it's better. on benchmarks"
They seem to just say "Here is our new model it's better, but you figure out how"
Maybe I'm old now.
Chinese models are still better. The EU and US need better open-source models.
Pretty sweet. I wonder if they can continue to distill this down so it runs smaller. I think we're at a point where more intelligence isn't really needed for a lot of use cases.
What is the point of releasing "The Most Capable US Open‑Weight Model?" I can understand why I might care about the location of a model serving provider, but why should I as a model user care about the location that an open weights model was trained?
It would be nice if the other players in the market would do the same and release yesteryears models as something to run at home.
Especially for all the people that got emotionally attached to specific checkpoints and their.. uh.. "character".
It's really nice that all AI pages are using canvas animations. A few days ago, Gemma 4's animation wasn't that impressive. Same with Muse. So far, the best animations I've seen are from OpenAI and Anthropic.
Ehm.. I have a model that runs on my older gaming desktop that scores as well (or better) as this model does on DeepSWE. This model does better than Flash Next in many areas and I am sure is much stronger overall, but hmm.
Zuck said by himself that they will release the weights of Muse Spark 1.2, and now we only get that one. At least they bench similarly bad for the size on Artificial Analysis
Who would have thought that one of the top American tech companies would boast about how their latest product is the best in class in the US, not the world, because they're behind China.
It did feel very odd to read that.
Prices need to come down on openrouter because right now it's a model that's not as good as deepseek v4.1 flash and it costs about 10x.
It performs worse than Muse Spark 1.3 and 1.2. And it has an outdated attention mechanism which at 1M tokens will need over 200GB just for the KV cache.
Very disappointing release. We're lucky to have Chinese models.
Glad that Meta is back on the open weights train in force.
How is this open weights? Open weights means anyone can use the weights. The license forbids certain categories.
I run a separate verification process, so I don’t encounter many errors while coding. However, I do have concerns regarding privacy.
Whats their incentive to share an open weight model? Heck, whats anyone?
In my poor enviroment, I can't use any model of open AI, so I wanna use it in opencode etc
Given exposed benchmaxxing of several past models from Meta,
I don't trust any benchmark numbers from Meta anymore.
They completely overhauled their team since then.
Muse wonder is still benchmaxxed.
Except the CEO.
“We believe AI is for everyone” *who can afford two M5 Studios with a total of 1 terabyte of memory.
Very big LLMs can still be useful even when you run them only at the speed of around 1 token per second that can be reached with almost any mini-PC or laptop with the weights streamed from SSDs.
While at that speed they will not be your main LLM, there are tasks that are still useful even when requiring many hours, e.g. an overnight run, for instance for checking what has been produced after many interactive iterations of a fast but small model.
Why aren't they using more modern benchmarks? Terminalbench 2.1 is old right?
What is the main driver for anyone to give away their hard earned weights? What is the catch?
Pissing off Dario, probably?
Or rather capturing the people Anthropic keeps alienating.
The Zuck has (in a different context) said that a few billions wasted are a decent bet (or something along those lines), so I think they're playing extra long games here with no real focus on profit any time soon.
To try to prevent as many people as possible from using one model, that gets all the data, and winner takes all.
spoiling the leaders' profit margins
Is it easy to use Muse to remove all the kinds of crap I don’t want in my Facebook feed?
Asking for a friend.
This is not an open model. License is not open license, and only weights are provided.
What kind of hardware does one need to run this model? Like a small datacenter node?
UD-IQ1_S is 180Gb. UD-Q4_K_XL is 510Gb. Still beyond the average prosumer.
I need this for my M6 Mini.
Good enough small models are the way.
Nice a Mastodon shoutout.
It's not open!!!
How so?
Feels dumber than GLM 5.3 (only open model I use extensively). But is really fast (fireworks.ai).
Only checked a single, large prompt that already ran today on claude and glm though..
Only if NVIDIA didn't shit on consumers, maybe we could run this with a single card today. But they are not even selling the 5090 32gb anymore. I hate this progress. Models are out, but there's no hardware to use them.
Why is Gemma not in the comparison table?
Probably too small, the comparisons are all larger models.
from what i understand its somewhere between muse spark 1.3 and 1.2 in some coding benches.
Never meta.
Do not run, do not support, do not use.
I wonder if all the AI companies are making dossiers on their users.
I would say meta and google definitely are. xAI? Probaly. The others? Perhaps. Perhaps Anthropic less than the others?
It appears to already be far beyond that, they are working with each other[0]:
>ChatGPT, Claude, Grok, DeepSeek, Perplexity, Gemini, Copilot, Le Chat de Mistral and Meta AI filter your data to Google, Meta, and several of the main data brokers.
[0] https://jorgegarciaherrero.com/en/prompt-like-a-butterfly-st...
https://jorgegarciaherrero.com/wp-content/interactivos/20260... [PDF]
Muse Spark Open IS MAGA!
great, I've been reading series of papers related to open source/open weight LLM Backdoors.
It's 100% possible, curious when will be the first incident.
Having something built by entity with HQ in the wester world is always a bit safer. That could be mistral, but they decided to be a consultancy company, not an ai lab.
Right. The US has famously never installed government-mandated backdoors in security-critical technology like Dual_EC_DRBG.
everyone is doing backdoors, except eu, they just receive bribes from those who install backdoors.
the only difference is: the backdoor is made to kill people like you or not. I'd rather have a backdoor that is installed to kill someone on the different continent, e.g Iran leadrship, not in my country/city.
2 replies →
Mistral just released Mistral Large 4?
and it's already not competitive from day 1
1 reply →
Link to papers?
can we stop these safety risk.... Feels like gpt 2/3 days, and now we know there wasn't any safety issue with gpt 2/3....
woahh man, thanks for the articles reference.
in what way is this "open"?
I tried Muse Spark 1.3 for a while because it's so cheap, but it's definitely in a lower league than Astra, Opus 5.5 and Argon. Still produces slop if you try and vibe code with it, whereas the more capable models write good code.
Try Muse Wonder Full, it's comparable.
Ah interesting. It's also comparable in price though!
So: WAR!
I hate it that zuck is the one promoting open weight models.
Good luck running that on your home infra lol, requires 25k in mac minis
odd...
Can we expect Cerebras will start offering it?
"Muse Spark Open was optimized to run on Mac Studio with M5 Ultra, and together they’re the best way to run a powerful model on your own desktop."
”Serve the model on either two Mac Studios, with M5 Ultra and 512GB of unified memory; or four Mac Studios with 256GB of unified memory, for a total of 1TB of unified memory, enough to hold a NVFP4 quantized version of Muse Spark Open”
Can not find any info on inference speed.
its around 20tps
Read it again. Two M5s.
Wait read that again. Two 512GB, unreleased model M5s.
This is gonna cost like 50k to run at home.
Closer to ~$30K if I had to guess. A 256 version is $11,299.00 (2TB SSD, 256GB RAM, 36-core CPU, 80‑core GPU), so yeah, 4 of those would be close to $50K but the 96->256 RAM upgrade is $4K, if we assume 512 will be double that it's ~$15K/ea.
But yes, still a massive amount of money and that's before we know the TPS or any other stats.
1 reply →
How much does the engineers at Meta make? Spend enough time together with others making $250k+ and eventually you start suggesting that two 512 GB M5s is something you can put on your desktop. I mean, it's one banana, Michael. What could it cost? 10 dollars?
Why would anyone want to run an LLM on Apple's hardware? Why not buy an NVIDIA GPU?
1 reply →
okay that's gonna be interesting to try
need to qualify with open and US. Lol!
Yawn.
[dead]
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[dead]
[dead]
[dead]
But does it play doom? /s
PS: I might need to do a doom benchmark, that’d be fun https://playground.jeffyclassify.com/#doom
Best I can do is p(doom)
Is that combination of tokens still relevant or did it leave the news cycle by now?
[dead]
[dead]
Tech people really are drug addicts. Rant about Zuck and Meta for 364 days a year and then jump over each other to be first in line the moment he shows something shiny.
[flagged]
Zuck is the good guy here. Damn
You don't know what is hiding in the model ...
By finding another way to siphon user data?
You trust anything he says?
US OPEN WEIGHT IS NOT DEAD THANKS ZUCK
Nemotron, Inkling, Gemma, Trinity Large, Snowflake Arctic were all released/updated in the last 6 months, and there are countless other smaller models out there. US open weight is very far from "dead", and certainly doesn't need Zuck as a savior.
Also IBM Granite has a small thinking model now (though it is slightly exasperating in use)
Is Google not gonna make another Gemma or something?
Even if they stopped, the Chinese model release hit an all-time high this summer so I think we're good.
Google doesn’t release big chungus models like this
As an aside I really wish people would stop naming their new software projects “Spark” given how many things already use that name.
The sad thing is I think it's like Elara, it's just a favorite LLM name for things and there's likely not a human behind it.
You guys be sniffing each others butts too deeply sometimes.