Comment by amanzi
8 hours ago
No mention of the performance of the models? I'm able to load a bunch of different models on my little mini-PC with 16GB RAM, but the performance is terrible. I always wonder what performance people are getting with local models that they find is acceptable?
Qwen3.8:27b-mlx on 64GB MBP M4, I can get up to 42tok/s, more often than not in the ~30 range.
i have a 512gb ram m3 ultra mac studio setup with a gas city that runs one of my companies. today was the first time ever that a local model (GLM5.3 8-bit) was able to match fable5 in our tests.
GLM-5.3-Flash at true 8-bit: 341 GB on disk, 328 GB resident, 288 experts across 46 layers, loads in 65 seconds. • 18.7 tokens/s generation, 35 tokens/s prompt, on a desk, on a $0 per-token bill. • Runs beside our whole agent city on one box with ~130 GB to spare. • Review test: caught 6 of 6 planted P1 defects, zero false positives, same score as the frontier model we pay for. • CRM test: 11 of 11 required records extracted, zero wrong writes, 45 minutes, first local model to clear the bar. • Serving a 131k-token window today; the model itself supports 1,048,576. Widened to 4 concurrent slots and still have 50gb+ of excess ram.
granted my cto still isn't moving all of our inference to glm5.3 but we've identified 40%+ that is currently handled by fable that we're routing locally instead and will do concurrent requests to verify/compare responses for a while.
I run a similar setup to the one he described on similar hardware. I run bifrost and llama swap though (tailscale rocks). My local model usage is for some out of band batch processing one of my personal apps uses. Basically a personalized recommender for media, it curates stuff for me based on a database i've compiled over years, so non-interactive. For that use case, I don't really care that it might take a few minutes to run. It's free. The machine is just sitting there anyway. I have tried using qwen-coder and opencode on my M5 Max 128gb and compared to claude code it's painful. I did setup a workflow where claude plans, qwen executes (unattended overnight, again b/c it's slow) and then claude reviews. I benchmarked this several times and I ended up using MORE tokens with claude because it had to 'fix' all the qwen issues. While the code it produced was 'good enough' the fixes were worth it so I just stick to coding task using API models (codex and claude).
> It's free
It isn't, the cost is included in your electricity bill, not even talking about the cost of your time to set it up. It's very possible that it costs you more than a cloud mode would, you just don't want to calculate it properly.
> It's very possible that it costs you more than a cloud mode would
...which is almost always true in a single request/reply mode and never true in batch mode. Single request usually 2x-3x more expensive than cloud and batch mode 2x-3x cheaper. Now, for narrow tasks, a finetuned tiny 8b model would dramatically outperform SOTA frontiers for a fraction of price, esp. on energy efficient hardware like Apple.
Can you share a bit more about your bifrost and llama swap setup? I’m facing memory constraints and am looking for a managed model solution that will help with hot swapping loaded models and stay-warm concurrency. Ideally with prioritization.
What do you want to know? Just start llama-swap with the models i have downloaded, add llama-swap as a provider in bifrost, expose the models you want and they become available in one single endpoint you can use in anything like opencode, openwebui or anything that speaks openai.
I have an M4 pro (48 GB ram) and I run Gemma 4 26b a4b at 52 tok/s and Qwen 3.5b a3b at 72 tok/s. Both 4bit quantized. These are enough for my needs and the performance is more than good enough. I'm not running the MLX version of the Gemma model, if I did the inference speed would likely be a bit better. I wouldn't use them for coding features though.
> enough for my needs
Which are...?
Some examples (keep in mind this is all indefinitely free for me, no burning quota away):
1. Getting information (such as information about hardware unfamiliar to me) when not connected to the internet, which happens occasionally in my case.
2. Continuing to learn Rust by way of toy examples, puzzles, and comparing aspects of various solutions, for example from LeetCode.
3. Reformatting data, for example from a PDF to a markdown table, or converting receipt images to text.
4. Simple translation/explanation (e.g. I'm teaching my wife one of the languages I speak but sometimes may not know/have the words to explain the full nuance of a translated word).
5. Summarization. One of the webnovels I'm reading has some very boring parts I don't want to slog through, in those cases I simply make the LLM summarize that part and move on.
Etc., you get the idea. It's not unusable for coding, but it would make many mistakes when making a whole feature and the context lengths are limited to around 30k-40k tokens by my RAM. I could give it access to the web but I simply use an online model when I need that sort of thing, again partly due to the context limit.
Edit: The MLX version of Gemma 4 26b a4b does about 62 tok/s.
Are you using the right configuration for your own CPU?
On a Laptop with 32 GB RAM and Iris Xe integrated graphic card, I get between 11-18 Tokens/Second with Qwen 3.8 27B and llama.cpp with sysl Intel optimisations. Same results with the vulkan back end, although sometimes it ends in weird segmentation faults due to the memory consumption.
> I'm able to load a bunch of different models on my little mini-PC with 16GB RAM, but the performance is terrible
With all due respect, I'm not clear why you are so surprised ?
By your own admission its a little mini-PC with 16GB RAM, I'm not sure what miracles you were expecting ?
Its a bit like complaining Rasperry Pi performance is terrible when trying to compile the Linux kernel.
I can't imagine using CPU... Oh I did twice.
If you are work from home and do dishes between prompts you can get a gpt3-like result.
I found it useful when I was... Well I didn't find it useful. But an Nvidia 3060 let me ask unethical questions pretty fast.
With which model. I have a 3060 with a bunch of system ram
It’s not. Do it as a hobby or for privacy but for performance just use a frontier model api. You’re paying less than cost for something that would take tens of thousands to set up locally.
That's not even remotely close to being true, even once you account for capex. You have to look at the actual usage, look at the token limits. Even if you're paying Anthropic $200k/month for scale-tier, you're going to blow through your token limits trying to run max output 24/7. Three users running Opus 4.8 at max non-stop will probably clean your monthly allowance from daddy Dario in less than a week.
With an 8x MI355x cluster at full tilt and including cooling, your power draw runs ~17kW. That's what it looks like when it's running full tilt. To be fair, hey that's pretty expensive. It does mean 8 multi-trillion parameter models unquantized running 24/7 without pause. And you get the full month like that, your monthly token limit is the time in a month. That cluster, the electrical upgrade, the cooling setup, and the electricity to run it all costs less in 2 months than your maximum affordance from Anthropic does in the same time period. Two billing cycles, and realistically it's more like two weeks. In 4 quarters you've wasted over a million. Like, what are we talking about here?
Now if you aren't using AI all that much, which is perfectly valid, and especially if you aren't using it at its absolute maximum, the story changes. Because even though at that point you're not paying nearly as much in electricity to run the cluster anymore, you still have the $300k+ capex to get the setup in the first place. But if we're not redlining it non-stop, then we're not really talking about performance anymore, are we? If your org never comes close to hitting token limits, it's probably because AI is rather marginal for you. Which again, is perfectly valid. I don't even use AI professionally.
Fact of the matter is, if your corp can justify the capex for a cluster and makes heavy use of AI, you are literally burning money by not having one in your building. The numbers are painfully obvious. Even deepseek isn't as cheap. This is before we get into things like LoRAs, custom inference pipelines, etc. which you know are kind of important if you actually care about model performance.
Here’s an experiment: purchase an anthropic pro max subscription for $200/m. Now go buy the hardware to run DeepSeek’s equivalent. In a year, who spent more?
4 replies →
> With an 8x MI355x cluster at full tilt and including cooling, your power draw runs ~17kW. That's what it looks like when it's running full tilt. To be fair, hey that's pretty expensive.
Pretty expensive is an understatement. You couldn’t buy one of these if you wanted to right now. If you could it would be multiple hundreds of thousands of dollars.
> It does mean 8 multi-trillion parameter models unquantized running 24/7 without pause
You can’t even run one unquantized multi-trillion parameter (>=2T) model on 8 x MI355x with enough context for concurrent users. I don’t know how you think it’s going to run 8 of them at the same time. Did you mean 8 concurrent sessions?
Your math is way off across this post. If replacing an Anthropic subscription for a whole company was as easy as buying a box for the office and then breaking even in 2 months, it wouldn’t be some little secret that we only discover in a comment online.
14 replies →
What're you using that monster for?
It does make me wonder how the hosted stuff is so cheap. For pretty much everything else, hosted/rented is more expensive but offers better convenience and flexibility. But for AI, even if you consider the total lifetime cost and are utilizing it heavily. You never break even by buying.
They're not cheap at all. I did one xhigh Qwen 3.8 27B agentic coding task last week via OpenRouter and it cost me like $10.
99% of the cost was in input tokens, I only used like 100k ish output tokens. It was a one shot task asking the agent to implement proxy injection to Guice. It did a pretty amazing job.
If you were to use hosted LLMs for a lot of agentic coding, a maxed out M5 Ultra Mac Studio would pay for itself in under a year.
2 replies →
It's a time sharing agreement, just like old-school mainframes and such. You're not getting a full machine to yourself, but a few cycles at a time.
There are economies of scale but there’s also a data center bubble (probably) so there might be some selling dollars for fifty cents going on.
I watched someone at a fortune 20 company get embarrassed for buying a Mac to run a 70B model in 2025.
He was a lead engineer, so after he announced it wasn't going to work, everyone pretended it never happened. But we all knew.
Sounds like a really rude workplace. Who cares if he wants to try running things locally?
These are estimates: https://tokenstead.ai/find/results?hardware_id=23&use_case=c...
I honestly wouldn’t bother with local models right now unless I either had a 5090 and was happy with running Qwen 3.8 27B, or a pair of DGX Sparks running DSv4 flash, or better, 2x6000 RTX Blackwells. Those are the kinds of rigs that the local model enthusiasts are running. With the GPU setups, you’re looking at generally >100tps generation in single stream, and >10k tps of prefill, so it’s snappier than Claude code, which somewhat makes up for it being dumber.
That said, it is really cool to be able to run an LLM on eg a Mac laptop. Just not a better experience on almost any metric for interactive use than eg Claude Code, beside privacy and guardrails.
>I honestly wouldn’t bother with local models right now unless I either had a 5090 and was happy with running Qwen 3.8 27B
How's the actual performance of Qwen 3.8 27B? On deepswe it supposedly performs slightly worse than gpt 5.6 luna high[1], but I can't help but think they've been benchmaxxed.
[1] https://deepswe.datacurve.ai/, https://unsloth.ai/docs/models/qwen3.8#benchmarks
Not sure, I haven't run it, I've just been running DS V4 Flash non-stop since it came out, and that's replaced a lot of my Claude Code usage. People seem very impressed, though, it seems like it trades vram/world knowledge for extra thinking time, which I think is a good trade for local. tbf, I've heard luna's not great at coding. Fast and good for things like classifiers, summarization, though.
A friend and I were actually discussing today how benches show Luna Max at about par on coding with Sol Medium, but how it's nowhere near in reality. We were speculating that maybe it's because a lot of benches are best-of-n, and should probably be worst-of-n, because variance in performance is killer with large coding projects. Consistency is what lets you actually build on this stuff.
Keep in mind these downloadable models use 3-10x the amount of tokens as well. You really can’t beat a couple $20 subscriptions.
https://quesma.com/benchmarks/babaisbench/
1 reply →