As an ex Meta employee, this is a little sad but not massively surprising. 'Number go up' is the core performance evaluation metric until PSC is done and you move on.
Why are resource limits considered at all aside from models accidentally fork bombing themselves?
I thought the benchmark was supposed to be about terminal use and specifically chaining together lots of bash tool calls. Which test cases does this matter for?
Terminal bench 2 isn't simply about 'somehow' getting a task done, it intends to measure real world behavior of an agent, including environment awareness in a given situation.
A few examples from memory:
1. This task [1] asks the agent to train a CNN under 1 CPU, 2GB RAM, 10GB storage. If you allow high resources, weaker models often succeed (the most clock time actually goes in waiting for the network to train).
2. This task [2] asks agents to implement a complete MIPS interpreter in JavaScript in 1 cpu and 2GB RAM. A common failure mode is OOM, at least in the earlier buggy versions that models run to get feedback. When OOM hits, the task is killed, no do-overs.
3. A lot of tasks involve building projects with a single core supplied. If you use -j12 type options, it will actually be _slower_ to build and the task will more likely miss the timeout. Having more threads squeezes the end to end time. This is a big one actually since the most common failure mode (from what I have seen) is the task timeout hitting before the agent finishes
Thats what is wrong with close source models, we dont know exactly what we are paying for, a superior base model or a well thought harness for benchmaxing
This doesn't seem that big of a deal to me? I mean, in any other area where I want an assessment of a product, I'm not going to trust what the product producer says about it at face value -- obviously they're going to be biased. This is the whole raison d'etre for independent testing, like https://artificialanalysis.ai.
I get your point but I'm not sure it matters all that much.
Did harbor / tb2.1 cap the swap available to docker runs?
There used to be a bug that would allow dockerized instance runs to use more memory than the specs allowed. Some of the original tasks weren't really possible to complete without exploiting swap. Even the oracle solutions didn't pass if you stopped docker from having access to swap.
I think crack-7z-hash and filter-js-from-html had that problem off the top of my head, but i haven't looked at this in months, so i'm not sure
What that link describes is basically the motivation to go from terminal bench 2.0 to 2.1. The latter simply fixed the common issues/complaints. There is a long github discussion on tbench's about it
I had a few days of preview access, which was long enough to put together a plugin for LLM. You can try the model out in the terminal like this:
uv tool install llm
llm install llm-meta-ai
llm keys set meta-ai
# paste API key here
llm -m meta-ai/muse-spark-1.1 "Generate an SVG of a pelican riding a bicycle"
It does feel like a chore in weeks like this one where there are new models landing every day.
I'm also increasingly worried that I'm not providing enough value in my model reviews. It's really hard to get a useful and credible idea for the strengths of the new models.
Maybe Zuck should double down on his "spoiler" role with models rather than compete head-to-head.
He doesn't have to match Anthropic or OpenAI model revenue if he can deflate theirs by 99%.
All he has to do is keep spending a few billion dollars developing frontier models, release them as open weights, and turn coding models into a commodity. He also needs a good OSS reference harness to match. Very few people are in a position to do this and for it to make business sense.
That's quite likely where things are headed regardless, and he could speed it up significantly.
We should all hope models move from proprietary products to commodities the way compilers did.
This may be one of the best things Zuck could do for the world.
The goal is not for meta to take their market, the goal would be for meta to damage their competitors.
If meta releases an open-weight LLM that is not Chinese made, cheaper to run than the SOTA premiums, etc, it would lower the number of people paying for frontier labs models. We saw with with early LLAMA models, but they didn’t keep up in the race with v4.
Meta would benefit from this, not from increased revenue at the hands of open LLMs, but from reduced competition. Meta competes with Google for ad spend, and lowering the Google revenue (or increasing costs) from AI reduces the competitive advantage. OpenAI wants to build an ad engine, so same thing will apply there too - make it less-revenue-generating to compete. Meanwhile G, OpenAI, and Anthropic are huge talent sinks that they have to compete with, especially for ML talent which is core to Metas business goals (ads). Finally, Meta needs lots of GPUs to train their ad engine models. By reducing the revenue-per-GPU of these labs, they’re reducing demand on a core revenue generating supply they have to compete for.
Software has several unique properties on both ends of its production process that make assertions of progress based on the software use case invalid.
Software is easy to define as “working”: just run it. But - useful software requires an absolute truck worth of code - 100k loc before you’re talking about a real product, or else dozens of iterations of a toy you make for yourself before it’s useful enough to quit toying with and just use for what you wanted it for.
Sure, the success of software is hard to anticipate and what “good” UX is is hard to pin down - that’s not what I’m talking about. I’m talking just making the code and having no lint errors. That shit is a slog but it’s a slog with a clear goal amenable to hill climbing.
Through that lens software is mostly pattern matching. It’s very rare that an activity in software construction is out of distribution because even if the core of the thing is novel it needs a massive blanket of UI and a tech stack and a production environment to run in and observability and and and and. You get it I hope.
Meanwhile most work out there is a mess of undocumented, un-codifiable detail with no objective criteria for success, only a very wide gradient of “job well done” to “what is this garbage go and fix it”.
We are solving the easy parts of software and soon all that’s left will be the parts that are just like other work. And then we engineers will also be doing mostly squishy subjective judgment stuff.
all he has to do is that prove builing these inst hard anymore. because the whole moat these companies have is the perception that building models at frontier is really hard .
Yeah, this is most directly comparable to xAI Grok 4.5. In both cases, directionally "opus level intelligence for haiku prices" which is a really big deal for application developers who want to include models like this in their applications. I have been testing switching out haiku and sonnet for Grok 4.5, and may give this a try too (it is quite a bit cheaper, particularly for cached).
> Yeah, this is most directly comparable to xAI Grok 4.5.
Grok 4.5 has a relatively high $0.50 per 1M cached input token rate, compared to $0.15 on this model.
Grok 4.5 cached input costs the same as Opus 4.8 cached input, which is going to make it a lot more expensive to use for multi-turn coding than many would assume from the $2/$6 headline numbers they led with.
This is still ridiculously expensive imagine having to pay $10 for 100 search results on Google, thats essentially what this is.
I really dont see how anyone's willing spend more than $1.50 per mm output. Let alone $15-50. Does anyone actually pay for usage based billing as a consumer?
Interesting how the prevalent opinion until yesterday seems to have been that OpenAI & Anthropic are irreversibly ahead and now with xAI and Meta at least delivered something that's competitive with useful models and cheap too. Granted, the narrative that the two leading labs are ahead still holds with Fable (and perhaps an upcoming GPT6), but it's not as over as common knowledge by the opinion leaders would have us believe.
People misinterpreted Google being behind as Anthropic and OpenAi being really ahead, when it was really just Google falling behind the same way it did with Tensorflow, Angular and GCP.
> when it was really just Google falling behind the same way it did with Tensorflow, Angular and GCP
Not sure I agree. Angular fell behind in popularity but was (is? unsure atm) still eminently usable. I gave gemini a test drive recently and it was horrendous, as in "picking dirt cheap Chinese model over gemini any day" bad, and with overzealous guardrails to boot. 3.1 pro feels a year behind and is extremely lazy. 3.5 flash feels like a model you’d run on your 128gb macbook, not something that was released a month ago and which costs a fair bit when used through api.
In any case: as of right now I think that we went from a three horse race to anthropic / openai as premium choices vs whatever is the Chinese fotm for a fraction of the cost. 3.5 pro better be a miracle if google wants to hang out with the big boys, otherwise their only strategy is hoping that both US labs go broke and they remain the last man standing.
> Interesting how the prevalent opinion until yesterday seems to have been that OpenAI & Anthropic are irreversibly ahead
Not the way you're implying?
The GLM 5.2 hype was blowing way before this. Neither xAI nor Meta have really made a difference in a different way - similar results / similar pricing (to GLM 5.2).
Where is the data retention policy information for paid API per-token uses? Every other provider has one and makes it clear how they handle your data. A quick look doesn’t show one for this new offering.
Yeah, I think it is definitely great. Having said that, I am still debating in my mind whether the volume of software engineers needed in the AI era is going to increase or decrease because of all of these advancements.
On the one hand, because it is easy to build products, more and more people will build. And more and more products and features will be built. However, a lot of people who are non-technical will also try to build, but they get stuck, and then they will need engineers. The sheer volume of product built by both experienced technical companies and non-technical novice startups and founders and wannabe founders is going to be massive. That is the bull case for having more software engineers needed in the near future.
On the other hand, in a year or so, people will build all these products, and most of them won't be able to market them, sell them and make money. Eventually, there won't really be a need for that many software engineers.
I think overall the bull case is probably going to win net net.
I see some similarities to 3D printing here. It’s great that everyone can make their own toothbrush holder (or whatever) but I’m probably not going to pay for someone’s weekend project.
I’m “seeing” more devs stepping into the SendCutSend stage where they’re cleaning up/fixing/productizing vibe coded projects so maybe there will be some new demand in that space?
The big thing to me is why are we even running these models on top of an operating system?
What I really want is Claude as a deep part of the operating system.
If that happens then a whole lot of the abstraction of software vanishes along with what we think of today as software jobs. I think many new forms of knowledge work would emerge from this though.
I would think that needs massive local compute but I can't imagine that is not the future down the line.
> On the one hand, because it is easy to build products, more and more people will build.
And those people won't need to be software engineers.
> but they get stuck, and then they will need engineers
You've implicitly assumed here that the AI systems will always be worse than the average engineer. That is IMO myopic. I'm not sure that it's even true now let alone in the nebulous future.
He came to X to post about this instead of his very own meta threads. This just shows how much interested he is to make this thing big, and of course, the cost can stay bearable for us considering all of these cash burn that these companies are doing
> No wonder we still can’t get climate change under control
This is was historically a money issue, being green used to be wildly more expensive.
Now being green is cheaper, the limiting factor is how fast PV and batteries can be made or imported.
Recent reports of the sum of all US data centres currently in planning, has a power demand exceeding the (capacity-factor-adjusted!) global annual supply of new PV.
This would be less of a problem, but still a problem, if Trump wasn't trying to get in the way of anything green, or if the companies building data centres decided to also support factories to make more PV.
* Planned new demand: 300 GW; PV factory capacity ~ 600 GW nameplate, but the capacity factor is 14% so that's really 84 GW on average.
Just got it working with codex in a container! FYI I think there is a bug most others will run into at the Codex:Muse interface.
It's some kind of parsing or integration error due to what I think is codex not anticipating server-side tool calling and how meta treats those ids... first couple times running codex with muse, it would fail on its first non-web search call.
Got it fixed, not personally sold on the bespoke server-side tool calling and indefinite file storage yet, but also a very cool model that I'm enjoying using so far!
Debugging and diagnosis is very tool call heavy, whether that's grepping / transforming logs, calling out to profilers/tracers, or even just writing up incident reports.
Bug diagnostics is about being okay at coding but better at tooling.
Given a good diagnostic report, it can be handed to opus for the fix.
Opus is okay at writing reports, but it still regularly gets table widths wrong in typst documents, leaving the last column full of text but only a handful of characters wide.
I wonder if we'll start to see that pattern with every new release. Tool use likely changes rapidly, so the newest, rather than most intelligent, model may always have an edge.
What you mean.. The tools are all just invoking bash and terminal/cli cmds and http requests. Paradigms that have existed and stayed mostly unchanged for decades.
This sounds... kind of useless? Really good JSON or similar constrained decoder performance is interesting, but normal decoder > tool validator loop with good error message > tool retry is almost always able to get a tool to work second try, and input is cached so it's not expensive.
I missed the fact that Meta was developing and releasing closed-weights models... bummer. Would be great to see some more progress with American open-weights models.
This not being available on Openrouter really makes it hard to test. I was going to compare vs Grok 4.5 and GPT-5.6 Luna, but I don't want to deal with signing up for Meta for it unless it checks out. Please Meta make this available.
Nearly every model can be found on OpenRouter and used with a single key. Meta Spark is not among them, but Grok and almost every other model is. That's how I try models I don't already have an account for.
Interesting that neither meta nor xai chose to do open source given that they are both clearly behind Google, OpenAI and anthropic - and a serious us open source offering would give them a clear foothold.
I suspect they have a brand problem from their social media ties and shady histories. I personally will never use their models, plenty of better alternatives. I'm now exclusively on open weight models
Very strong pricing, cheaper than Grok 4.5, particularly the cached reads. We'll have to wait to see if it's actually worth using (it's not on OpenRouter yet).
Competition for cheaper and efficient models is a good thing, regardless of if you don't like SpaceX, Meta, etc. Especially from US based labs
I for one am really glad to get competitive models that will push the major labs to bring prices down. While Chinese open source labs are also great, unfortunately when it comes to US/Western political pressure it won't often have as much of a bearing on labs bringing prices down, especially for enterprises.
Also if these numbers are true, this is truly breaking ground finally for Meta.
Good to see Meta finally back to releasing something at least worth evaluating. And it sounds like they did at least a bit skate to where the puck is going by focusing on tool and computer use.
It's a high quality benchmark for sure, but it being public means it's at risk of leaking into the models (unintentionally or not), right? For that reason I prefer to look at the private ones, like: HLE, SimpleBench, Kagi, ARC-AGI.
Everyone has been loving to shit on the Alexander Wang acquisition but this seems legitimately impressive to me?
Meta's AI org when from a total mismanaged dumpster fire for multiple years to delivering a competitive model in less than a year on essentially their first try?
As far as i remember, the entire AI org was essentially gutted and replaced with whoever Wang wanted to hire, and tbh that org completely failed to train llama 4 and I honestly doubt whatever techniques they used to ship llama 3 are at all relevant now. That was before reasoning models and the heavy emphasis on RL/post-training.
so yeah, this is essentially their first try with a completely new org.
Yes and Zuck effectively disbanded the entire team that did that. Not saying we shouldn't cast a critical eye on it, but it probably does warrant a second chance.
It's great that we have yet another competing models. The more models we have, the less likely we are subject to the ideologies and the controls thereof by the cults like Anthropic. And of course, it drives down the cost of tokens.
Lot more details in the linked report https://www.tbench.ai/leaderboard/terminal-bench/2.1
As an ex Meta employee, this is a little sad but not massively surprising. 'Number go up' is the core performance evaluation metric until PSC is done and you move on.
Why are resource limits considered at all aside from models accidentally fork bombing themselves?
I thought the benchmark was supposed to be about terminal use and specifically chaining together lots of bash tool calls. Which test cases does this matter for?
Terminal bench 2 isn't simply about 'somehow' getting a task done, it intends to measure real world behavior of an agent, including environment awareness in a given situation.
A few examples from memory:
1. This task [1] asks the agent to train a CNN under 1 CPU, 2GB RAM, 10GB storage. If you allow high resources, weaker models often succeed (the most clock time actually goes in waiting for the network to train).
2. This task [2] asks agents to implement a complete MIPS interpreter in JavaScript in 1 cpu and 2GB RAM. A common failure mode is OOM, at least in the earlier buggy versions that models run to get feedback. When OOM hits, the task is killed, no do-overs.
3. A lot of tasks involve building projects with a single core supplied. If you use -j12 type options, it will actually be _slower_ to build and the task will more likely miss the timeout. Having more threads squeezes the end to end time. This is a big one actually since the most common failure mode (from what I have seen) is the task timeout hitting before the agent finishes
[1] https://github.com/harbor-framework/terminal-bench-2-1/blob/...
[2] https://github.com/harbor-framework/terminal-bench-2-1/tree/...
> Which test cases does this matter for?
The test cases of "don't melt my computer" and "be a good (computational) neighbor"
Out of curiosity, how often are the resource limits the bottlenecks? What do harnesses do to help here - limit parallelism? More efficient tools?
The task could be verifiable in the environment so limiting its CPU and RAM could be to discourage brute forcing the answer.
Thats what is wrong with close source models, we dont know exactly what we are paying for, a superior base model or a well thought harness for benchmaxing
This doesn't seem that big of a deal to me? I mean, in any other area where I want an assessment of a product, I'm not going to trust what the product producer says about it at face value -- obviously they're going to be biased. This is the whole raison d'etre for independent testing, like https://artificialanalysis.ai.
I get your point but I'm not sure it matters all that much.
Did harbor / tb2.1 cap the swap available to docker runs?
There used to be a bug that would allow dockerized instance runs to use more memory than the specs allowed. Some of the original tasks weren't really possible to complete without exploiting swap. Even the oracle solutions didn't pass if you stopped docker from having access to swap.
I think crack-7z-hash and filter-js-from-html had that problem off the top of my head, but i haven't looked at this in months, so i'm not sure
Huh? What are you talking about?
https://www.anthropic.com/engineering/infrastructure-noise
Is anthropic benchmark maxxing and cheating on terminal bench too? They don't follow the strict resource "limits" either
What that link describes is basically the motivation to go from terminal bench 2.0 to 2.1. The latter simply fixed the common issues/complaints. There is a long github discussion on tbench's about it
1 reply →
I had a few days of preview access, which was long enough to put together a plugin for LLM. You can try the model out in the terminal like this:
Here's the result: https://tools.simonwillison.net/markdown-svg-renderer#url=ht...
For comparison, here's the pelican I got from Muse Spark 1: https://simonwillison.net/2026/Apr/8/muse-spark/
How do you find the time to “preview” so many models? It’s been a crazy time recently with the model releases. Does it ever feel like a chore?
It does feel like a chore in weeks like this one where there are new models landing every day.
I'm also increasingly worried that I'm not providing enough value in my model reviews. It's really hard to get a useful and credible idea for the strengths of the new models.
It's his job, and this is advertising (he is very good at native advertising on HN).
His primary interest nowadays is LLM's. Unsure if he gets paid for it, but he seems to take it as a dayjob.
1 reply →
Maybe Zuck should double down on his "spoiler" role with models rather than compete head-to-head.
He doesn't have to match Anthropic or OpenAI model revenue if he can deflate theirs by 99%.
All he has to do is keep spending a few billion dollars developing frontier models, release them as open weights, and turn coding models into a commodity. He also needs a good OSS reference harness to match. Very few people are in a position to do this and for it to make business sense.
That's quite likely where things are headed regardless, and he could speed it up significantly.
We should all hope models move from proprietary products to commodities the way compilers did.
This may be one of the best things Zuck could do for the world.
If he deflates their revenues, who is going to rent the compute from Meta?
The goal is not for meta to take their market, the goal would be for meta to damage their competitors.
If meta releases an open-weight LLM that is not Chinese made, cheaper to run than the SOTA premiums, etc, it would lower the number of people paying for frontier labs models. We saw with with early LLAMA models, but they didn’t keep up in the race with v4.
Meta would benefit from this, not from increased revenue at the hands of open LLMs, but from reduced competition. Meta competes with Google for ad spend, and lowering the Google revenue (or increasing costs) from AI reduces the competitive advantage. OpenAI wants to build an ad engine, so same thing will apply there too - make it less-revenue-generating to compete. Meanwhile G, OpenAI, and Anthropic are huge talent sinks that they have to compete with, especially for ML talent which is core to Metas business goals (ads). Finally, Meta needs lots of GPUs to train their ad engine models. By reducing the revenue-per-GPU of these labs, they’re reducing demand on a core revenue generating supply they have to compete for.
If it's good enough, they don't need to sell to Anthropic/OpenAI.
the way he could really be the spoiler king is to release an their training dataset to open source… doubt he’d go that far.
Coding models are not the destination. Coding models are just part of the bootstrapping process towards general intelligence.
Software has several unique properties on both ends of its production process that make assertions of progress based on the software use case invalid.
Software is easy to define as “working”: just run it. But - useful software requires an absolute truck worth of code - 100k loc before you’re talking about a real product, or else dozens of iterations of a toy you make for yourself before it’s useful enough to quit toying with and just use for what you wanted it for.
Sure, the success of software is hard to anticipate and what “good” UX is is hard to pin down - that’s not what I’m talking about. I’m talking just making the code and having no lint errors. That shit is a slog but it’s a slog with a clear goal amenable to hill climbing.
Through that lens software is mostly pattern matching. It’s very rare that an activity in software construction is out of distribution because even if the core of the thing is novel it needs a massive blanket of UI and a tech stack and a production environment to run in and observability and and and and. You get it I hope.
Meanwhile most work out there is a mess of undocumented, un-codifiable detail with no objective criteria for success, only a very wide gradient of “job well done” to “what is this garbage go and fix it”.
We are solving the easy parts of software and soon all that’s left will be the parts that are just like other work. And then we engineers will also be doing mostly squishy subjective judgment stuff.
He tried that with llama?
all he has to do is that prove builing these inst hard anymore. because the whole moat these companies have is the perception that building models at frontier is really hard .
Hitherto he's pretty much proved the opposite.
I guess we'll see how Meta did this time.
The pricing is insane: $1.25/$4.5 for 1M tokens, and $0.15 for cached input!
https://dev.meta.ai/docs/getting-started/pricing-rate-limits
Yeah, this is most directly comparable to xAI Grok 4.5. In both cases, directionally "opus level intelligence for haiku prices" which is a really big deal for application developers who want to include models like this in their applications. I have been testing switching out haiku and sonnet for Grok 4.5, and may give this a try too (it is quite a bit cheaper, particularly for cached).
> Yeah, this is most directly comparable to xAI Grok 4.5.
Grok 4.5 has a relatively high $0.50 per 1M cached input token rate, compared to $0.15 on this model.
Grok 4.5 cached input costs the same as Opus 4.8 cached input, which is going to make it a lot more expensive to use for multi-turn coding than many would assume from the $2/$6 headline numbers they led with.
1 reply →
The cached input pricing is a good ratio.
Compare with Grok 4.5 which came out at $2/$6 but then quietly charges $0.50 per 1M cached input tokens. That's as high as Opus 4.8!
Meta isn’t right now on the radar for most folks picking models.
If they have a really good model, it makes sense to subsidise it, to gain users, before they align prices with competitors.
this is not subsidizing. this is way too expensive for a no-name model.
2 replies →
Cheaper than Qwen 3.7 Max. Second indication, after Grok 4.5 ($2 in / $6 out), that the BigLabs are feeling the GLM 5.2 heat.
[flagged]
This is still ridiculously expensive imagine having to pay $10 for 100 search results on Google, thats essentially what this is.
I really dont see how anyone's willing spend more than $1.50 per mm output. Let alone $15-50. Does anyone actually pay for usage based billing as a consumer?
This is pretty cheap compared to anthropic opus and fable.
https://platform.claude.com/docs/en/about-claude/pricing
Model Base Input Tokens 5m Cache Writes 1h Cache Writes Cache Hits & Refreshes Output Tokens
Claude Fable 5 $10 / MTok $12.50 / MTok $20 / MTok $1 / MTok $50 / MTok
Claude Opus 4.8 $5 / MTok $6.25 / MTok $10 / MTok $0.50 / MTok $25 / MTok
Note Fable costs $50 MTok and Opus 4.8 costs $25 / MTok.
Yeah, DeepSeek V4 (Flash and Pro) is below $0.004 for 1M cache hits.
Even with usage based billing I'm below $1 writing code all day.
Interesting how the prevalent opinion until yesterday seems to have been that OpenAI & Anthropic are irreversibly ahead and now with xAI and Meta at least delivered something that's competitive with useful models and cheap too. Granted, the narrative that the two leading labs are ahead still holds with Fable (and perhaps an upcoming GPT6), but it's not as over as common knowledge by the opinion leaders would have us believe.
meh, models are mostly good enough now. Without a major breakthrough the only thing that matters now is cost.
People misinterpreted Google being behind as Anthropic and OpenAi being really ahead, when it was really just Google falling behind the same way it did with Tensorflow, Angular and GCP.
> when it was really just Google falling behind the same way it did with Tensorflow, Angular and GCP
Not sure I agree. Angular fell behind in popularity but was (is? unsure atm) still eminently usable. I gave gemini a test drive recently and it was horrendous, as in "picking dirt cheap Chinese model over gemini any day" bad, and with overzealous guardrails to boot. 3.1 pro feels a year behind and is extremely lazy. 3.5 flash feels like a model you’d run on your 128gb macbook, not something that was released a month ago and which costs a fair bit when used through api.
In any case: as of right now I think that we went from a three horse race to anthropic / openai as premium choices vs whatever is the Chinese fotm for a fraction of the cost. 3.5 pro better be a miracle if google wants to hang out with the big boys, otherwise their only strategy is hoping that both US labs go broke and they remain the last man standing.
1 reply →
> Interesting how the prevalent opinion until yesterday seems to have been that OpenAI & Anthropic are irreversibly ahead
Not the way you're implying?
The GLM 5.2 hype was blowing way before this. Neither xAI nor Meta have really made a difference in a different way - similar results / similar pricing (to GLM 5.2).
Where is the data retention policy information for paid API per-token uses? Every other provider has one and makes it clear how they handle your data. A quick look doesn’t show one for this new offering.
I personally do not like Meta, but I'll say this. The more competition, the better for regular consumers. (Enterprise too)
- Chinese models
- Grok
- Meta
- Google
- OpenAI
- Anthropic
I think this is a win. I'm building like crazy to take advantage of all these subsidized tokens while I can.
Meta's local llama models used to be the face of open source AI. The scene has really changed.
they likely got the Peter Theil newsletter proclaiming open source models are the antichrist
1 reply →
Yeah, I think it is definitely great. Having said that, I am still debating in my mind whether the volume of software engineers needed in the AI era is going to increase or decrease because of all of these advancements.
On the one hand, because it is easy to build products, more and more people will build. And more and more products and features will be built. However, a lot of people who are non-technical will also try to build, but they get stuck, and then they will need engineers. The sheer volume of product built by both experienced technical companies and non-technical novice startups and founders and wannabe founders is going to be massive. That is the bull case for having more software engineers needed in the near future.
On the other hand, in a year or so, people will build all these products, and most of them won't be able to market them, sell them and make money. Eventually, there won't really be a need for that many software engineers.
I think overall the bull case is probably going to win net net.
I see some similarities to 3D printing here. It’s great that everyone can make their own toothbrush holder (or whatever) but I’m probably not going to pay for someone’s weekend project.
I’m “seeing” more devs stepping into the SendCutSend stage where they’re cleaning up/fixing/productizing vibe coded projects so maybe there will be some new demand in that space?
3 replies →
The big thing to me is why are we even running these models on top of an operating system?
What I really want is Claude as a deep part of the operating system.
If that happens then a whole lot of the abstraction of software vanishes along with what we think of today as software jobs. I think many new forms of knowledge work would emerge from this though.
I would think that needs massive local compute but I can't imagine that is not the future down the line.
1 reply →
> On the one hand, because it is easy to build products, more and more people will build.
And those people won't need to be software engineers.
> but they get stuck, and then they will need engineers
You've implicitly assumed here that the AI systems will always be worse than the average engineer. That is IMO myopic. I'm not sure that it's even true now let alone in the nebulous future.
11 replies →
At least in China a lot of software developers are now struggling.
I think for a lot of type of software we have now reached peak employment.
Someone payed a few k just for a normal website.
5 replies →
To expand on Chinese models:
- DeepSeek
- GLM (Z.ai)
- Minimax
- Kimi (Moonshot)
- Hy3 (Tencent)
- Qwen (Alibaba)
(Each one of these with weights available to download and run locally)
- MiMo (Xiaomi)
GLM 5.2 is great, but is so rate limited now I no longer recommend it
6 replies →
He came to X to post about this instead of his very own meta threads. This just shows how much interested he is to make this thing big, and of course, the cost can stay bearable for us considering all of these cash burn that these companies are doing
Its the biggest technology race we have ever seen. Richest companies, smartest people, richest countries.
I do not know if competition is good, we will see in a few years.
Looking forward having a physical job for a change :D
A bit much describing our tech leadership as smartest people we've ever seen.
2 replies →
While data centers are still using lots of energy created from fossil fuels and many still evaporate water for cooling?
No wonder we still can’t get climate change under control
> No wonder we still can’t get climate change under control
This is was historically a money issue, being green used to be wildly more expensive.
Now being green is cheaper, the limiting factor is how fast PV and batteries can be made or imported.
Recent reports of the sum of all US data centres currently in planning, has a power demand exceeding the (capacity-factor-adjusted!) global annual supply of new PV.
This would be less of a problem, but still a problem, if Trump wasn't trying to get in the way of anything green, or if the companies building data centres decided to also support factories to make more PV.
* Planned new demand: 300 GW; PV factory capacity ~ 600 GW nameplate, but the capacity factor is 14% so that's really 84 GW on average.
How is every company able to show itself at the top of every benchmark?
First look what models are worse in a set of self selected benchmarks.
Second, compare to older versions of competitor s models.
Still does not look good? Compare to own previous models.
Not much moat, incremental improvements, cherry picking models to compare.
To be fair, seems more correct to compare against similar strength models if your main edge is pricing.
At this point comparing to Gemini is a free Bingo space.
Anyone deep in the AI realm know which is the gold standard benchmark for coding?
Wait to the exact moment your model is ahead on at least N benchmarks then publish.
They're being called "trust me bro benchmarks" for a reason ( ･ั ﹏ ･ั )
Just got it working with codex in a container! FYI I think there is a bug most others will run into at the Codex:Muse interface.
It's some kind of parsing or integration error due to what I think is codex not anticipating server-side tool calling and how meta treats those ids... first couple times running codex with muse, it would fail on its first non-web search call.
Got it fixed, not personally sold on the bespoke server-side tool calling and indefinite file storage yet, but also a very cool model that I'm enjoying using so far!
https://github.com/accretional/awesome-muse-spark/blob/main/...
Their published benchmarks seem to indicate that it's pretty good at coding and multimodal, but VERY good at successful tool calls.
What kind of use case would be best for that shape?
Debugging and diagnosis is very tool call heavy, whether that's grepping / transforming logs, calling out to profilers/tracers, or even just writing up incident reports.
Bug diagnostics is about being okay at coding but better at tooling.
Given a good diagnostic report, it can be handed to opus for the fix.
Opus is okay at writing reports, but it still regularly gets table widths wrong in typst documents, leaving the last column full of text but only a handful of characters wide.
Gemini 3.5 flash is better than fable at tool calling. Tool calling is probably one of the easier things to do post training for.
I wonder if we'll start to see that pattern with every new release. Tool use likely changes rapidly, so the newest, rather than most intelligent, model may always have an edge.
What you mean.. The tools are all just invoking bash and terminal/cli cmds and http requests. Paradigms that have existed and stayed mostly unchanged for decades.
1 reply →
This sounds... kind of useless? Really good JSON or similar constrained decoder performance is interesting, but normal decoder > tool validator loop with good error message > tool retry is almost always able to get a tool to work second try, and input is cached so it's not expensive.
Things are not always that simple, eg https://lucumr.pocoo.org/2026/7/4/better-models-worse-tools/
The avg coding session has hundreds or thousands of tool calls. Even a 5% failure rate noticeably notches up token use and cost. See Gemini.
1 reply →
It seems to trade blows with GPT 5.5 and Opus 4.8 in performance while being cheaper than GLM 5.2.
Glad to see Meta back on track! Users will benefit greatly from this competition.
> Model API is not available in your region.
:(
Well, Vietnam is not in the list of restricted territories.
Anyway, what is "your region" ?
Is this where I am now, or is it where I activated my Oculus 2 five years ago ?
Same in Argentina. It's almost surely a region whitelist for now (it's the only reason Argentina ever gets blocked).
Can’t you just use VPN?
Your Meta account profile region has to be in the U.S., and they do Geo IP and SMS verification.
Same in Canada. lol
I missed the fact that Meta was developing and releasing closed-weights models... bummer. Would be great to see some more progress with American open-weights models.
This not being available on Openrouter really makes it hard to test. I was going to compare vs Grok 4.5 and GPT-5.6 Luna, but I don't want to deal with signing up for Meta for it unless it checks out. Please Meta make this available.
Despite Muse being relatively average, I've actually used the Meta AI webchat LLM since it released.
The reason: Its writing style feels "unique", and I find it pleasant to read for science-based topics.
I never ask _ONLY_ Meta AI, but the answer it gives is almost always in a distinctly different style than other frontier LLM's.
I think this is because of the unique JEPA architecture they have, but that's a layman's hunch.
I think it does not use JEPA.
Why are the plans and pricing for all these products so complicated.
I don't know where I need to sign up to try it out. What is pricing? Is it API or subscription, what?
I had the exact same experience with Grok 4.5 as well.
Nearly every model can be found on OpenRouter and used with a single key. Meta Spark is not among them, but Grok and almost every other model is. That's how I try models I don't already have an account for.
Interesting that neither meta nor xai chose to do open source given that they are both clearly behind Google, OpenAI and anthropic - and a serious us open source offering would give them a clear foothold.
Open source would make them an instant credible leader, major fumble (still can be fixed)
I suspect they have a brand problem from their social media ties and shady histories. I personally will never use their models, plenty of better alternatives. I'm now exclusively on open weight models
Very strong pricing, cheaper than Grok 4.5, particularly the cached reads. We'll have to wait to see if it's actually worth using (it's not on OpenRouter yet).
That's what one does when its product and public perception is way behind competitors.
Since it was making rounds yesterday here is what muse generated for the 3d rubik's cube prompt - https://ashish01.github.io/rubik-muse/
Hold my beer...
https://chat.z.ai/space/t19sx5kvw631-art
How are people trying this? I don't see it on openrouter. Any ways of testing this without subscribing to meta stuff?
Probably need to wait some hours/1-2 days and openrouter will add it.
Thanks. I was asking because I couldn't find even their previous 1.0 model there.
Yeah, no thanks. I cannot think of a worse company to trust with additional personal data.
Me neither, though LLMs also provide services that don’t involve personal or sensitive data
Competition for cheaper and efficient models is a good thing, regardless of if you don't like SpaceX, Meta, etc. Especially from US based labs
I for one am really glad to get competitive models that will push the major labs to bring prices down. While Chinese open source labs are also great, unfortunately when it comes to US/Western political pressure it won't often have as much of a bearing on labs bringing prices down, especially for enterprises.
Also if these numbers are true, this is truly breaking ground finally for Meta.
There are US companies hosting open weight models for enterprise, we just enabled Fireworks.ai for the devs
Good to see Meta finally back to releasing something at least worth evaluating. And it sounds like they did at least a bit skate to where the puck is going by focusing on tool and computer use.
Considering the DeepSWE result (imho if you're gonna give value to benchmarks this is one of the best) it's not good enough.
It's a high quality benchmark for sure, but it being public means it's at risk of leaking into the models (unintentionally or not), right? For that reason I prefer to look at the private ones, like: HLE, SimpleBench, Kagi, ARC-AGI.
Everyone has been loving to shit on the Alexander Wang acquisition but this seems legitimately impressive to me?
Meta's AI org when from a total mismanaged dumpster fire for multiple years to delivering a competitive model in less than a year on essentially their first try?
Not their first try. There’s been reporting about how they’ve kept pushing their model releases back because of underwhelming performance.
... i dont think internal iteration counts dude. thats just called in-development.
How is it their first try? They were leading the race with Llama 3.x a few years ago.
As far as i remember, the entire AI org was essentially gutted and replaced with whoever Wang wanted to hire, and tbh that org completely failed to train llama 4 and I honestly doubt whatever techniques they used to ship llama 3 are at all relevant now. That was before reasoning models and the heavy emphasis on RL/post-training.
so yeah, this is essentially their first try with a completely new org.
1 reply →
They were leading the race in a niche category a few years ago. Now they are, according to some benchmarks, even on the right playing field.
I'm still confused is it available to public via some sort of subscription?
My trust factor is gone with Meta right now. Has there been any independent analysis to confirm they didn't cheat on benchmarks again?
They cheated again: https://news.ycombinator.com/item?id=48847019
No they didn't. Please read: https://www.anthropic.com/engineering/infrastructure-noise
1 reply →
Is this the model trained on Meta "draftees"? Are we seeing this in the jump on JobBench?
Meta is back in the game, albeit not at the top. Impressive stuff, nonetheless.
Weren't they caught multiple times gaming the benchmark even more so then the rest?
Yes and Zuck effectively disbanded the entire team that did that. Not saying we shouldn't cast a critical eye on it, but it probably does warrant a second chance.
1 reply →
Let me assure you, literally everybody does this
1 reply →
They are not open source anymore, right?
Is building models on the frontier really easy now or something?
haha no thx
Tried to get access to the API, apparently the model API is not available in my region...
I have questions regarding if I should even care but I don't so Meta please keep enjoying the irrelevance. lmao
Right, amazing because for me also... "My region" being Canada.
I'm going to assume the only "region" that's permitted is the USA.
Not opensource.
Haha their demo is AI spamming restaurants on Instagram. This is going to go really well.
A lot of these benchmarks are unfamiliar. Are labs just choosing the ones that make them look best?
This is not open-weights, right?
Correct
It's great that we have yet another competing models. The more models we have, the less likely we are subject to the ideologies and the controls thereof by the cults like Anthropic. And of course, it drives down the cost of tokens.
[dead]
Still no weights? Useless