Note that this is not the only way to run Qwen 3.5 397B on consumer devices, there are excellent ~2.5 BPW quants available that make it viable for 128G devices.
I've had great success (~20 t/s) running it on a M1 Ultra with room for 256k context. Here are some lm-evaluation-harness results I ran against it:
The method in this link is already using a 2-bit quant. They also reduced the number of experts per token from 10 to 4 which is another layer of quality degradation.
In my experience the 2-bit quants can produce output to short prompts that makes sense but they aren’t useful for doing work with longer sessions.
This project couldn’t even get useful JSON out of the model because it can’t produce the right token for quotes:
> *2-bit quantization produces \name\ instead of "name" in JSON output, making tool calling unreliable.
I can't say anything about the OP method, but I already tested the smol-IQ2_XS quant (which has 2.46 BPW) with the pi harness. I did not do a very long session because token generation and prompt processing gets very slow, but I think I worked for up to ~70k context and it maintained a lot of coherence in the session. IIRC the GPQA diamond is supposed to exercise long chains of thought and it scored exceptionally well with 82% (the original BF16 official number is 88%: https://huggingface.co/Qwen/Qwen3.5-397B-A17B).
Note that not all quants are the same at a certain BPW. The smol-IQ2_XS quant I linked is pretty dynamic, with some tensors having q8_0 type, some q6_k and some q4_k (while the majority is iq2_xs). In my testing, this smol-IQ2_XS quant is the best available at this BPW range.
Eventually I might try a more practical eval such as terminal bench.
This is always the problem with the 2-bit and even 3-bit quants: They look promising in short sessions but then you try to do real work and realize they’re a waste of time.
Running a smaller dense model like 27B produces better results than 2-bit quants of larger models in my experience.
> This is always the problem with the 2-bit and even 3-bit quants: They look promising in short sessions but then you try to do real work and realize they’re a waste of time.
It would be nice to see a scientific assessment of that statement.
4 bit is as low as I like to go. There are KLD and perplexity tests that compare quantizations where you can see the curve of degradation, but perplexity and KLD numbers can be misleading compared to real world use where small errors compound over long sessions.
In my anecdotal experience I’ve been happier with Q6 and dealing with the tradeoffs that come with it over Q4 for Qwen3.5 27B.
Generally the perplexity charts indicate that quality drops significantly below 4-bit, so in that sense 4-bit is the sweet spot if you're resource constrained.
What's the tok/s you get these days? Does it actually work well when you use more of that context?
By the way, it's been a long time since I last saw your username. You're the guy who launched Neovim! Boy what a success. Definitely the Kickstarter/Bountysource I've been a tiny part of that had the best outcome. I use it every day.
So it starts at 20 tps tg and 190 tps pp with empty context and ends at 8 tps tg and 40 tps pp with 250k prefill.
I suspect that there are still a lot of optimizations to be implemented for Qwen 3.5 on llama.cpp, wouldn't be surprised to reach 25 tps in a few months.
> You're the guy who launched Neovim!
That's me ;D
> I use it every day.
So do I for the past 12 years! Though I admit in the past year I greatly reduced the amount of code I write by hand :/
Apologies to others for the offtopic comment, but thank you so much for neovim. I started using Vim 25 years ago and I almost don't know how to type without a proper Vi-based editor. I don't write as much code these days, but I write other stuff (which definitely needs to be mostly hand written) in neovim and I feel so grateful that this tool is still receiving love and getting new updates.
I don't think MLX supports similar 2-bit quants, so I never tried 397B with MLX.
However I did try 4-bit MLX with other Qwen 3.5 models and yes it is significantly faster. I still prefer llama.cpp due to it being a one in all package:
- SOTA dynamic quants (especially ik_llama.cpp)
- amazing web ui with MCP support
- anthropic/openai compatible endpoints (means it can be used with virtually any harness)
- JSON constrained output which basically ensures tool call correctness.
- routing mode
Yes. Note that the only reason I acquired this device was to run LLMs, so I can dedicate its whole RAM to it. Probably not viable for a 128G device where you are actively using for other things.
Reading the details, he is using 2-bit quantization and reduced the number of experts per token from 10 down to 4 to get 5 tokens/sec. Cool proof of concept but it’s far from the quality and performance of the 397B model as normally used. Dropping the number of experts is particularly misleading.
This is some interesting work, but applying such extreme measures to LLMs to get them to run severely degrades quality. I know he claims negligible quality loss, but in my experience 2-bit quantizations are completely useless for real work. You can get them to respond to prompts, but they lose their intelligence and will go around in circles.
He also shows 5-6 tokens per second. Again that’s impressive for a large model on limited hardware but it’s very slow. Between the severely degraded model abilities and the extremely slow output the 397B result should be considered an attempt at proving something can technically run, not evidence that it can run well and produce output you’d expect from a 397B model.
He even mentions the obvious problems with his changes:
> *2-bit quantization produces \name\ instead of "name" in JSON output, making tool calling unreliable.
So right out of the gate this isn’t useful if you want to do anything with it. He could have tried smaller models or less quantizations to get actual useful output from the model, but it wouldn’t look as impressive. It’s honestly getting kind of exhausting to read all of these AI-coded (admitted in the link) and AI-written papers made more for resume building. It would have been interesting to see this work applied to running a useful model that hadn’t been lobotomized instead of applying tricks to get an impressive headline but useless output.
This is not an ordinary LLM benchmark, it's streaming experts' weights from storage. It opens up running very large (near-SOTA, potentially SOTA) MoE models on very limited hardware, since you no longer need enough RAM for the entirety of the model's parameters. The comparison to 20 t/s local AI models is simply not fair.
> *2-bit quantization produces \name\ instead of "name" in JSON output, making tool calling unreliable.
I was wondering about that statement. Shouldn't it restrict sampling to only tokens that produce valid JSON matching the schema during a tool call? On the other hand, I have heard a lot about how even production LLM providers don't always call tools accurately, so I suppose either it's hard to implement what I described or there's something I haven't thought of that makes it impossible.
To be honest, I'm getting tired of a "laptop" in every one of these clickbait titles turning out to be $3000 Macbook. Sure, it's impressive to achieve this degree of the LLM compression, but I really don't like that the title implies local LLM becomes a viable for an average person with the actual hardware being out of reach for 99%.
I enjoy them because I have a reasonably beefy laptop (non-Mac, though, so I can't try out this particular project), and it's nice to see what you can do with laptops in this space.
I think, though, maybe consider your biases? Whenever I see a headline like this, I absolutely don't assume it's going to run on any random laptop. I don't even expect it to run well, or at all, on my laptop (for stuff that isn't Mac-only, that is), which has an iGPU. I'm generally a big LLM/AI skeptic, but I find your brand of cynicism/dismissal to be kinda boring and uninteresting.
You can probably go lower than $3000, I expect, ime, an M1 max with 64GB ram to have similar performance, and you can find such one used with less than $2000 or so probably.
In any case, I do not think that the range of people that have an M1/2/3/4 max macbook is that narrow, eg people who may do video editing or who benefit from having one of the fastest multicore laptops. It is handy to be able to do work with a machine you already may own for separate reasons, though it is definitely more to the side of a "pro device" than "basic consumer device".
I ran full version of this model without any swapping on cluster of 2x $3000 laptops (strix halo zbook 128GB) at about 20 tokens per second.
I would say it is in reach for normal person. If anything buying it was great investment, it is work tool, I will probably sell it for more than what I bought it for :)
..and to think $2=3k is basically a minimum to really use local LLMs effectively, with prices definitely rising globally. Most people will just stick with online services for LLMs.
The github page mentions that a naïve mmap approach is bottlenecked by per-page overhead. Can this be mitigated by setting up explicit "huge" pages? (2M using the CONT PTE feature if the "native" page size is 16k; 32M using a PMD level block mapping; or 1G using the CONT PMD feature.) Does macOS support this out of the box? Alternatively, one might use a simple mmap and then something like posix_fadvise to set up prefetching of the data.
The quality degradation at 2-bit is a real issue. For actual work tasks, a well-tuned 30B at 4-bit usually outperforms a 70B+ at 2-bit in my experience. The expert reduction on top of that compounds things - you're essentially running a fairly different model. Still interesting to see the upper bound of what consumer hardware can attempt, even if the result isn't production-ready.
Very impressive! I wonder if there is a similar path for Linux using system memory instead of SSD? Hell, maybe even a case for the return of some kind of ROMs of weights?
Loading experts to system memory is supported by most local-AI frameworks. But you do not gain much by running that part of the decode on GPU, since decode is not compute-limited and the CPU-GPU transfer involves overhead. It's best to use the GPU for speeding up the shared part of the model.
Using system memory and CPU compute for some of the layers that don’t fit into GPU memory is already supported by common tools.
It’s workable for mixture of experts models but the performance falls off a cliff as soon as the model overflows out of the GPU and into system RAM. There is another performance cliff when the model has to be fetched from disk on every pass.
It's less of a "performance falls off a cliff" problem and more of a "once you offload to RAM/storage, your bottleneck is the RAM/storage and basically everything else no longer matters". This means if you know you're going to be relying on heavy offload, you stop optimizing for e.g. lots of VRAM and GPU compute since that doesn't matter. That saves resources that you can use for scaling out.
It depends on the model and the mix. For some MoE models lately it’s been reasonably fast to offload part of the processing to CPU. The speed of the GPU still contributes a lot as long as it’s not too small of a relative portion of compute.
My thoughts exactly. Something like this could make it so that modest GPU capacity, like a pair of 3090s , and lots of RAM could make big inference more practical for personal labs
This project shows an interesting automated search for engineering problems that I like to see more.
The experience of utilizing tiered storage (gpu vram, ram, and ssd) is generally poor for a lot of LLM inference engines out there, e.g., llama.cpp, sglang, vllm, etc..
My own experience shows that both weight and KV cache offload to ram on sglang and vllm is unavailable or unusable. Copying extra parameters from documents and adding them to already working commands results in errors. Llama.cpp does support weight offload, but the experience is not pleasant, low pcie (gpu <-> ram) utilization, low gpu utilization, and really low tokens per second.
2 bit quantization lobotomizes the model but is impressive nonetheless! Maybe one day we'll be able to have intelligent 2 bit quants... I wonder.
my version supports - 4bit quantization, hybrid streaming (Disk + ram), arbitrary model compatibility, tested on Mamba2, and lets up the framework for LM Studio integration
I leveraged this work (Credit to Danveloper) and am in the middle of making this work on more practical models and quants. It still uses flash streaming, but done so with a control knob so you can choose how much ram and how little ram to use. In the craziest case, it uses as little ram as possible but is very slow, however, in the balanced case you use some ram and it's much faster.
I designed it around the intelligence dense Nemotron 3 Nano 30B and Nemotron Cascade 2 30B models (which are smaller, more intelligence density) and can run on low end 16GB machines, though you can run arbitrarily large models on larger machines (designed for very low end, but capable of high end).
This is a very impressive result. If I understand correctly the bottleneck is the SSD in this architecture - the author seems to get almost 15GB/s - but I seem to remember the max b/w was about 8GB/s. What am I missing?
IO is very bursty in these setups. When the router results are in you can start loading experts from SSD. In this brief moment the SSD is saturated.
Outside of that the SSD is idling.
Table 3 shows for K=4 experts an IO of 943 MB/Tok at 3.15 Tok/s giving an average IO of 2970 MB/s far below what the SSD could do.
I'm not sure, but not all expert weights are used immediately. Maybe they could do async reads for the down tensors parallelizing compute with IO.
Not sure if this works on Mac, I only tested my larger than RAM setup on Linux with io_uring O_DIRECT reads and I saw that about 20% of total reads do finish while my fused upgate matmul is already running.
The github page mentions that you can't overlap SSD traffic and GPU compute on Apple Silicon, you get heavy contention for the shared hardware resources.
The finding about removing the 9.8 GB Metal LRU cache for a 38% speedup is the most interesting part. Same lesson as PostgreSQL's advice against application-level buffer pools that compete with the OS page cache : the hardware memory compressor doing 130K decompressions/sec was pure overhead.
Curious about the remaining gap: 5.7 tok/s vs 18.6 theoretical (from SSD bandwidth). Is the ~70% overhead mostly GPU compute on non-expert layers (attention, norm), or is there I/O scheduling room left?
Does this mean that it should be possible to load up a system with ~10 (seems to me at least the number of active experts) SSDs to get 40 tok/s even on truly gigantic models?
SSD bandwidth will ultimately be limited by the amount of PCIe lanes you have available (for something other than the Apple Silicon internal storage). So the approach has inherent limitations. You can of course scale out to multiple systems to get more throughput.
You can use this approach with Intel Optane, which is wearout-resistant unlike NAND and can thus substitute for RAM. Last I checked, it was available quite cheap on the secondary market, ~$1/GB as opposed to ~$15/GB or more for DRAM. (Of course that's nowhere near as cheap as NAND, which is around ~$0.1/GB but quite wearout-prone with heavy writes.)
Yeah, PCIe is the bottleneck. The point being that whether the data originates from RAM or from NVME or Optane, you cannot get data to the GPU faster with RAM than with SSDs.
Meanwhile PCIe switches exist. So why not build:
1 CPU + memory + ...
N PCIe switch with each 1 low-memory GPU + 6 NVME drives (in theory 5 can saturate the GPU)
Each of those should only bother the CPU when they have some tokens produced and have plenty of PCIe lanes to get at their data.
Such a setup should be able to get a 6 to 8 times speedup from the solution detailed here, and a model compute increase should make relatively little difference in performance.
I'm not sure how you're going to get 10 SSDs to fit inside a laptop. And if you're not going to use a laptop, then you might as well get an expensive machine with plenty of RAM, no? Even with RAM prices as crazy as they are these days, that's probably still cheaper than a machine with enough PCIe bandwidth/lanes/ports for 10 SSDs, not to mention the cost of the SSDs themselves.
Though I guess this could be an interesting area of research, regardless of the cost.
4.4 tok/s with reliable structured output is a solid local benchmark altho the question is whether SSD streaming introduces per-token latency variance that messes up tool call parsing downstream. The gap between 400 GB/s unified memory bandwidth and 17.5 GB/s SSD reads means you're in the hot path pretty much every time an expert isn't cached.
If you don’t care about turnaround time you can do that.
Most LLM use cases are about accelerating workflows. If you have to wait all night for a response and then possibly discover that it took the wrong direction, misunderstood your intent, or your prompt was missing some key information then you have to start over.
I don’t let LLMs write my code but I do a lot of codebase exploration, review, and throwaway prototyping. I have hundreds to maybe thousands of turns in the LLM conservation each day. If I had to wait 10X or 100X as long then it wouldn’t be useful. I’d be more productive ignoring a slow LLM and doing it all myself.
> If you have to wait all night for a response and then possibly discover that it took the wrong direction, misunderstood your intent, or your prompt was missing some key information then you have to start over.
If you have to wait overnight because the model is offloading to disk, that's a model you wouldn't have been able to run otherwise without very expensive hardware. You haven't really lost anything. If anything, it's even easier to check on what a model is doing during a partial inference or agentic workload if the inference process is slower.
"If you have to wait all night for a response and then possibly discover that it took the wrong direction, misunderstood your intent, or your prompt was missing some key information then you have to start over."
This exact problem exist for rendering, when you realize that after a long render an object was missing in the background and the costly frame is now useless. To counter that you make multiple "draft" renders first to make sure everything is in the frame and your parameters are properly tuned.
There's definitely use cases for this for long running tasks, like doing research, but for typical use cases they require way too much constant supervision and interaction
The author states that the code was written by Opus, and afaik AI-written code is not considered copyrightable. Without copyright on the code, there should not be something prohibiting you from running it or even redistributing it. Of course this may come down to the extent of human contribution.
fwiw, the opposite view was recently taken for node js. GH hides most of the conversation by default due to size, but the gist is that a new VFS feature is being proposed for node that was largely written using an LLM. I think I've more often seen the view these days that if you used, steered, and likely modified code generated by an AI, you generated the code and hold the copyright.
...you can't redistribute code without a license, but surely you can legally run it, can't you?
Like, if I write a blog post and put it on my blog, you're allowed to read it, right?
Heck, if my blog contains some Javascript code I wrote, I would imagine your web browser is allowed to run that code without opening you up to copyright infringement, even if I didn't provide an explicit license.
Really interesting approach. Curious how the 2-bit quantization affects the model's reasoning ability on longer chains of thought vs shorter prompts. The benchmarkslook solid but real-world usage seems like a different story based on the comments here.
48 GB is not consumer hardware. But fundamentally, there are economies of scale due to batching, power distribution, better utilization etc.., that means data center tokens will be cheaper. Also, as the cost of training (frontier) models increases, it's not clear the Chinese companies will continue open sourcing them. Notice for example, that Qwen-Max is not open source.
Nothing obviously prevents using this approach, e.g. for 3B-active or 10B-active models, which do run on consumer hardware. I'd love to see how the 3B performs with this on the MacBook Neo, for example. More relevantly, data-center scale tokens are only cheaper for the specific type of tokens data centers sell. If you're willing to wait long enough for your inferences (and your overall volume is low enough that you can afford this) you can use approaches like OP's (offloading read-only data to storage) to handle inference on low-performing, slow "edge" devices.
It is consumer hardware in the sense that Macbook Pros come with this RAM size as base and that you can buy them as a consumer, without having to sign a special B2B contract, show that your company is big and reputable enough, and order a minimum of 10 or 100.
Technically that's correct (which as we all know is the best kind of correct), but really, how many consumers are buying a high-end MacBook Pro with 48GB or more of RAM? That's a very small percentage of the population. In these kinds of discussions, "consumer" is being used as a proxy for "something your average home laptop buyer might have". And a 48GB MBP is not that.
I know it's annoying, because a 48GB MBP is indeed technically "consumer hardware", but please understand the context and don't be pedantic. You know what the GP meant. (And if not, that's... kinda on you.)
Assuming 'moat' – they'll push the frontier forward; they don't really have to worry until progress levels off.
At that point, I suppose there's still paid harnesses (people have always paid for IDEs despite FOSS options) partly for mindshare, and they could use expertise & compute capacity to provide application-specific training for enterprises that need it.
0.1 GB per full-attention layer and "The model has 60 transformer layers: 45 GatedDeltaNet (linear attention) + 15 standard full attention." So, 1.5 GB.
impressive, i wish someone takes a stab at using this technique on mobile gpu's even if it does not use storage it would still be a win. I am running llama.cpp on adreno 830 with oepncl and i am getting pathetic 2-3t/s for output tokens
I honestly don't get "why" despite having done similar things myself, e.g. run on a model on a VR headset itself.
I mean I've done it because I could, so I imagine others are doing that too. But then... once it's done I don't actually use it. I ticked that box but eventually when STOA aren't that useful I have a hard time imagining actual positive use cases (... not like offline spam or naughty chat in the woods) that would benefit from such technically impressive demos.
This reads like an unfairly shallow dismissal. 4.4 tok/s is more than par for the course (even acknowledging that the model was heavily quantized and stripped down by limiting active experts) for something that's streaming weights from storage in order to run on vastly more limited hardware than the model was designed for. It's even more impressive since the OP writeup mentions Apple Silicon hardware cannot overlap GPU compute with SSD access.
The problem is that the hardware is still like $3000. Making anything run on Macs is an exercise in futility. And its a shame that people get duped into buying Macs for LLM inference.
$3000 for running a 397B total parameters model is quite a bargain. The Mac is being used for its access to fast internal storage here since that's the key bottleneck, you could probably achieve similar outcomes with conventional (even fairly low-end) iGPU/APU hardware plus a fast PCIe x4 5.0 SSD (which would also allow you to overlap SSD transfers with iGPU/APU compute), but the cost would also be in a similar range. (Unless you carefully chose low-end e.g. Intel hardware with proper PCIe x4 5.0 NVMe support - which is still quite uncommon, especially for laptops.)
can you post the final result (or as far as you got before you killed it) to show us how cohesive and good it is? I'd like to see an example of the output of this.
The original tokens have Ġ instead of space. I had this issue too when writing an inference engine for Qwen. You have to "normalize" those special characters.
The technical write up is great, but Mac users should not get too excited just yet on running 300B+ parameter models locally as the TPS isn't that good.
>...at 4.4+ tokens/second
That is even when it is using 4-bit quantization and it is still at that speed.
> The entire 209GB model streams from SSD through a custom Metal compute pipeline.
This is my main problem.
If I were to run this on a Mac SSD, 24/7 for heavy usage such as Openclaw, that is going to significantly reduce the lifetime of the SSD.
Can't imagine using this in the long term right now, but improvements will follow. Still a great write up anyways.
> If I were to run this on a Mac SSD, 24/7 for heavy usage such as Openclaw, that is going to significantly reduce the lifetime of the SSD.
How sure are you about that? I've never looked closer at how a large LLM with mixture of experts architecture switches between expert modules, but staying on roughly the same topic for the use (as it often would when editing the same codebase), I wouldn't be surprised to see the switches of composition are fairly rare, fairly small, and to the extent it happens it's repeated reads from the flash disk rather than writes it tends to cause.
Afaik the experts are not usually very interpretable, and generally would be surprised if at least one does not change every token. I don't know what happens in practice, but I know at least during training, nothing is done to minimize the number of expert switches between tokens.
I'd have thought at least a tiny explicit penalty term for switching, to discourage messing around with the composition without any expected gains from it.
If one is to use these on hardware that can't keep everything loaded I guess someone should examine how it works out in practice. Interpretability may be be a too much to ask, but I can't spontaneously see any reason why the experts can't at least be pushed to incorporate what's needed to remain the good choice for a longer segment.
The switching is done by layer, not just per token. Every layer is loading completely different parameters, you don't really benefit from continuity. You're generally better off shifting this work to the CPU, since CPU RAM is more abundant than the GPU's VRAM hence it matters less that so much of it is "wasted" on inactive expert layers. Disk storage is even more relatively abundant, so offloading experts to disk if you can't keep them in RAM (as OP does) is the next step.
Eh. I mean, 4 tokens a second works fine if you're patient. Go do something else while you wait.
I feel like whenever I'm trying to find information on which local models will work on my hardware, I have to overestimate because people don't know how to wait for things.
I've had great success (~20 t/s) running it on a M1 Ultra with room for 256k context. Here are some lm-evaluation-harness results I ran against it:
More details of my experience:- https://huggingface.co/ubergarm/Qwen3.5-397B-A17B-GGUF/discu...
- https://huggingface.co/ubergarm/Qwen3.5-397B-A17B-GGUF/discu...
- https://gist.github.com/simonw/67c754bbc0bc609a6caedee16fef8...
Overall an excellent model to have for offline inference.
In my experience the 2-bit quants can produce output to short prompts that makes sense but they aren’t useful for doing work with longer sessions.
This project couldn’t even get useful JSON out of the model because it can’t produce the right token for quotes:
> *2-bit quantization produces \name\ instead of "name" in JSON output, making tool calling unreliable.
Note that not all quants are the same at a certain BPW. The smol-IQ2_XS quant I linked is pretty dynamic, with some tensors having q8_0 type, some q6_k and some q4_k (while the majority is iq2_xs). In my testing, this smol-IQ2_XS quant is the best available at this BPW range.
Eventually I might try a more practical eval such as terminal bench.
This is always the problem with the 2-bit and even 3-bit quants: They look promising in short sessions but then you try to do real work and realize they’re a waste of time.
Running a smaller dense model like 27B produces better results than 2-bit quants of larger models in my experience.
It would be nice to see a scientific assessment of that statement.
In my anecdotal experience I’ve been happier with Q6 and dealing with the tradeoffs that come with it over Q4 for Qwen3.5 27B.
They did reduce the number of experts, so maybe that was it?
By the way, it's been a long time since I last saw your username. You're the guy who launched Neovim! Boy what a success. Definitely the Kickstarter/Bountysource I've been a tiny part of that had the best outcome. I use it every day.
I ran llama-bench a couple of weeks ago when there was a big speed improvement on llama.cpp (https://github.com/ggml-org/llama.cpp/pull/20361#issuecommen...):
So it starts at 20 tps tg and 190 tps pp with empty context and ends at 8 tps tg and 40 tps pp with 250k prefill.I suspect that there are still a lot of optimizations to be implemented for Qwen 3.5 on llama.cpp, wouldn't be surprised to reach 25 tps in a few months.
> You're the guy who launched Neovim!
That's me ;D
> I use it every day.
So do I for the past 12 years! Though I admit in the past year I greatly reduced the amount of code I write by hand :/
@justinmk deserves the credit for this!
Have you compared against MLX? Sometimes I’m getting much faster responses but it feels like the quality is worse (eg tool calls not working, etc)
I don't think MLX supports similar 2-bit quants, so I never tried 397B with MLX.
However I did try 4-bit MLX with other Qwen 3.5 models and yes it is significantly faster. I still prefer llama.cpp due to it being a one in all package:
- SOTA dynamic quants (especially ik_llama.cpp) - amazing web ui with MCP support - anthropic/openai compatible endpoints (means it can be used with virtually any harness) - JSON constrained output which basically ensures tool call correctness. - routing mode
Update: I just did a quick asitop test while inferencing and the GPU power was averaging at 53.55
This is some interesting work, but applying such extreme measures to LLMs to get them to run severely degrades quality. I know he claims negligible quality loss, but in my experience 2-bit quantizations are completely useless for real work. You can get them to respond to prompts, but they lose their intelligence and will go around in circles.
He also shows 5-6 tokens per second. Again that’s impressive for a large model on limited hardware but it’s very slow. Between the severely degraded model abilities and the extremely slow output the 397B result should be considered an attempt at proving something can technically run, not evidence that it can run well and produce output you’d expect from a 397B model.
He even mentions the obvious problems with his changes:
> *2-bit quantization produces \name\ instead of "name" in JSON output, making tool calling unreliable.
So right out of the gate this isn’t useful if you want to do anything with it. He could have tried smaller models or less quantizations to get actual useful output from the model, but it wouldn’t look as impressive. It’s honestly getting kind of exhausting to read all of these AI-coded (admitted in the link) and AI-written papers made more for resume building. It would have been interesting to see this work applied to running a useful model that hadn’t been lobotomized instead of applying tricks to get an impressive headline but useless output.
6t/s is unbearable, if you used it with OpenCode you would be waiting 20+ minutes per turn.
At 1t/s you are never going to get anything done. At 6t/s, it's significantly degrades the experience, one mistake setting you back 20-30 minutes.
At ~20t/s it's much more usable.
I was wondering about that statement. Shouldn't it restrict sampling to only tokens that produce valid JSON matching the schema during a tool call? On the other hand, I have heard a lot about how even production LLM providers don't always call tools accurately, so I suppose either it's hard to implement what I described or there's something I haven't thought of that makes it impossible.
I think, though, maybe consider your biases? Whenever I see a headline like this, I absolutely don't assume it's going to run on any random laptop. I don't even expect it to run well, or at all, on my laptop (for stuff that isn't Mac-only, that is), which has an iGPU. I'm generally a big LLM/AI skeptic, but I find your brand of cynicism/dismissal to be kinda boring and uninteresting.
In any case, I do not think that the range of people that have an M1/2/3/4 max macbook is that narrow, eg people who may do video editing or who benefit from having one of the fastest multicore laptops. It is handy to be able to do work with a machine you already may own for separate reasons, though it is definitely more to the side of a "pro device" than "basic consumer device".
I would say it is in reach for normal person. If anything buying it was great investment, it is work tool, I will probably sell it for more than what I bought it for :)
That is a very wealthy country centric thing to say.
Even with a MoE model, which has to move a relatively small portion of the weights around, you do end up quite bandwidth constrained though.
It’s workable for mixture of experts models but the performance falls off a cliff as soon as the model overflows out of the GPU and into system RAM. There is another performance cliff when the model has to be fetched from disk on every pass.
I assumed the only reason this particular project wouldn't be usable on Linux is because it uses Metal...
Hand written... by GPT? ;)
The experience of utilizing tiered storage (gpu vram, ram, and ssd) is generally poor for a lot of LLM inference engines out there, e.g., llama.cpp, sglang, vllm, etc..
My own experience shows that both weight and KV cache offload to ram on sglang and vllm is unavailable or unusable. Copying extra parameters from documents and adding them to already working commands results in errors. Llama.cpp does support weight offload, but the experience is not pleasant, low pcie (gpu <-> ram) utilization, low gpu utilization, and really low tokens per second.
https://github.com/matt-k-wong/mlx-flash
2 bit quantization lobotomizes the model but is impressive nonetheless! Maybe one day we'll be able to have intelligent 2 bit quants... I wonder.
my version supports - 4bit quantization, hybrid streaming (Disk + ram), arbitrary model compatibility, tested on Mamba2, and lets up the framework for LM Studio integration
I leveraged this work (Credit to Danveloper) and am in the middle of making this work on more practical models and quants. It still uses flash streaming, but done so with a control knob so you can choose how much ram and how little ram to use. In the craziest case, it uses as little ram as possible but is very slow, however, in the balanced case you use some ram and it's much faster.
I designed it around the intelligence dense Nemotron 3 Nano 30B and Nemotron Cascade 2 30B models (which are smaller, more intelligence density) and can run on low end 16GB machines, though you can run arbitrarily large models on larger machines (designed for very low end, but capable of high end).
Outside of that the SSD is idling.
Table 3 shows for K=4 experts an IO of 943 MB/Tok at 3.15 Tok/s giving an average IO of 2970 MB/s far below what the SSD could do.
I'm not sure, but not all expert weights are used immediately. Maybe they could do async reads for the down tensors parallelizing compute with IO.
Not sure if this works on Mac, I only tested my larger than RAM setup on Linux with io_uring O_DIRECT reads and I saw that about 20% of total reads do finish while my fused upgate matmul is already running.
Edit: Typos
Curious about the remaining gap: 5.7 tok/s vs 18.6 theoretical (from SSD bandwidth). Is the ~70% overhead mostly GPU compute on non-expert layers (attention, norm), or is there I/O scheduling room left?
You can use this approach with Intel Optane, which is wearout-resistant unlike NAND and can thus substitute for RAM. Last I checked, it was available quite cheap on the secondary market, ~$1/GB as opposed to ~$15/GB or more for DRAM. (Of course that's nowhere near as cheap as NAND, which is around ~$0.1/GB but quite wearout-prone with heavy writes.)
Meanwhile PCIe switches exist. So why not build:
1 CPU + memory + ...
N PCIe switch with each 1 low-memory GPU + 6 NVME drives (in theory 5 can saturate the GPU)
Each of those should only bother the CPU when they have some tokens produced and have plenty of PCIe lanes to get at their data.
Such a setup should be able to get a 6 to 8 times speedup from the solution detailed here, and a model compute increase should make relatively little difference in performance.
Though I guess this could be an interesting area of research, regardless of the cost.
To render movies we happily wait for the computer to calculate how lights bounce around, for hours even days.
So why not do the same with AIs? Ask big question to big models and get the answer to the universe tomorrow?
Most LLM use cases are about accelerating workflows. If you have to wait all night for a response and then possibly discover that it took the wrong direction, misunderstood your intent, or your prompt was missing some key information then you have to start over.
I don’t let LLMs write my code but I do a lot of codebase exploration, review, and throwaway prototyping. I have hundreds to maybe thousands of turns in the LLM conservation each day. If I had to wait 10X or 100X as long then it wouldn’t be useful. I’d be more productive ignoring a slow LLM and doing it all myself.
If you have to wait overnight because the model is offloading to disk, that's a model you wouldn't have been able to run otherwise without very expensive hardware. You haven't really lost anything. If anything, it's even easier to check on what a model is doing during a partial inference or agentic workload if the inference process is slower.
This exact problem exist for rendering, when you realize that after a long render an object was missing in the background and the costly frame is now useless. To counter that you make multiple "draft" renders first to make sure everything is in the frame and your parameters are properly tuned.
https://github.com/nodejs/node/pull/61478
https://blog.platformatic.dev/why-nodejs-needs-a-virtual-fil...
Like, if I write a blog post and put it on my blog, you're allowed to read it, right?
Heck, if my blog contains some Javascript code I wrote, I would imagine your web browser is allowed to run that code without opening you up to copyright infringement, even if I didn't provide an explicit license.
So far ollama will use the 12GB and then give up
It’s a MacBook.
I know it's annoying, because a 48GB MBP is indeed technically "consumer hardware", but please understand the context and don't be pedantic. You know what the GP meant. (And if not, that's... kinda on you.)
The context is this is something I can pick up at an Apple Store and not some rig I have to build with NVIDIA cards.
I led with:
> get closer and closer to consumer hardware
I think this demonstrates getting closer, whether you think a MacBook is consumer hardware or not. But I'm the one being pedantic.
At that point, I suppose there's still paid harnesses (people have always paid for IDEs despite FOSS options) partly for mindshare, and they could use expertise & compute capacity to provide application-specific training for enterprises that need it.
here we go: https://huggingface.co/collections/trillionlabs/tri-series
Welp, I know where those tokens came from.
I mean I've done it because I could, so I imagine others are doing that too. But then... once it's done I don't actually use it. I ticked that box but eventually when STOA aren't that useful I have a hard time imagining actual positive use cases (... not like offline spam or naughty chat in the woods) that would benefit from such technically impressive demos.
anyone suggesting that it's a reasonable speed should find another career
You can buy a 3090 for $2k, and run QWEN3.5 at 50+ token a second, and it will do everything you need, especially if you give it enough context.
MacBook Pro M5 Pro (64GB RAM)
>...at 4.4+ tokens/second
That is even when it is using 4-bit quantization and it is still at that speed.
> The entire 209GB model streams from SSD through a custom Metal compute pipeline.
This is my main problem.
If I were to run this on a Mac SSD, 24/7 for heavy usage such as Openclaw, that is going to significantly reduce the lifetime of the SSD.
Can't imagine using this in the long term right now, but improvements will follow. Still a great write up anyways.
How sure are you about that? I've never looked closer at how a large LLM with mixture of experts architecture switches between expert modules, but staying on roughly the same topic for the use (as it often would when editing the same codebase), I wouldn't be surprised to see the switches of composition are fairly rare, fairly small, and to the extent it happens it's repeated reads from the flash disk rather than writes it tends to cause.
If one is to use these on hardware that can't keep everything loaded I guess someone should examine how it works out in practice. Interpretability may be be a too much to ask, but I can't spontaneously see any reason why the experts can't at least be pushed to incorporate what's needed to remain the good choice for a longer segment.
I feel like whenever I'm trying to find information on which local models will work on my hardware, I have to overestimate because people don't know how to wait for things.
Also, reading data doesn't cause SSD wear.