For anyone wondering “how slow is this?”

IIUC, Kimi K3 on RTX 6000 Ada (48GB) takes 292 s/token

https://github.com/lyogavin/airllm/releases/tag/v3.1.0

I hope I'm not the only one who misread it as 292 tokens/s and got excited momentarily

I do think these “run a bigger model than will fit in VRAM” projects are necessary steps, but are they functionally useful or helpful to anyone currently? For example, is anyone out there running a big Qwen for coding on a 16-32GB machine with these techniques?

> but are they functionally useful or helpful to anyone currently?

I've been building a SaaS that deals with data that can't be distributed to third parties. Some of the useful AI stuff I can add is not time sensitive and can run overnight. Things like this allow me to use higher quality models without selling my house for GPUs.

> but are they functionally useful or helpful to anyone currently?

Yes and no, depend on your expectations. Some/many like to run local LLMs just for the sake of it, so anything will do.

MoE are useful on PC systems, at the condition of having high enough memory bandwidth (and large amounts of RAM) - that is, Threadripper/Pro.

The advantage of MoE is that only a subset of the model's experts is used for each token, so not all weights need to be present in VRAM at once. The remaining weights can reside in system RAM, although moving and accessing them still carries a substantial performance cost (and that's why high memory bandwidth is needed).

Does MoE help with multimodality? Can it in general enable reasoning in imagery (technical drawings, diagrams, schematics) rather than text-based?

MoE has nothing to do with multimodality.

MoE is a concept proposed in 1991, before the deep learning era (which is before what I call the transformers era). You can think of it like sharing.

Contrary to popular belief; 'experts' in MoE LLMs do not specialize. There's no expert trained to be good at maths, or python, or writing, or whatever. It's an inference optimization.

As for reasoning in non-text modalities, you might find this paper interesting :) https://huggingface.co/papers/2502.05171

Wait I thought the router ends up specializing the experts?

Like there is no explicit goal aside from each 'expert' getting roughly equal weight?

And it happens that when you train the router you do end up passing certain classes of problem to each expert - just as a training result nothing as clean as a python expert. But math vs creative writing will tend to rely on different experts over the majority of the inference?

I do not know what I am talking about, this is my limited understanding...

> You can think of it like sharing.

was this meant to read "sharding"?

Are people getting decent tokens/second throughput? Some of these demos crawl at 1 tok/s or worse, which limits their utility.

yes, I average 80-120 tok/s on my RTX 3080 with gemma 4 and faster with Qwen 3.5. The main use-case here is just code-monkey agents. I'm not looking for architectural guidance, but an agent to take a spec and complete it.

And is this using conventional model loading (all in VRAM), or are you streaming it in some way?

in theory, QWEN3.6-35B-A3B should run as fast as a 3B model; and in practice, it can be just as dumb.

I don't need any of these to help with coding. skill issue.

If I could justify wear and tear and electricity, I was willing to do something like this for batch processing. The batches would be a bunch of prompts whose outputs I'd look at the next day. Maybe common operations, like QA or refactoring, on whatever software I wrote.

If so, I could use a larger model than I have real-time hardware for. The largest, well-trained models can often get the output mostly right in one try. I also would be using AI's as a supplement to, not replacement for, my own brain. So, issues with the outputs wouldn't be a problem because I'm just keeping what's helpful.

If I still need to re-generate it all, it might still save money over time by avoiding cloud costs. Also, hardware that's already paid for is a sunk cost that doesn't inflate over time. Glitches in loading or destroying VM's might blow up into a big bill.

at 5 minutes per token, you could look at the results next week

one week later: "It says 'You're absolutely right! Let me look at the seams so I'm checking, not guessing—' and I guess that's when my SSD melted."

Turns out the SSD was load-bearing.

At that point, how does this compare with simply running the model on the CPU?

It matches my coding speed...its ok.

Ahaha thank you, I naively assumed the unlabeled graph in the readme was tps, not spt!

Running 292 tps on K3 would make your gpu a money printer.

spt not tps

I wonder what this measures in J/token.

Assuming 30% gpu power utilization because of all the loading and unloading 29.2 kJ per token

Damn 15 AK47 bullets per token

How many is that in tokens per Scaramucci?

It is a bit more than four tokens per millifortnight.

that's 0.003 tokens/second. To get an hour's work done that's normally 30 tokens/second (108k output tokens in an hour) will take 416 days at this rate. And if you're using 100 watts, during that time you will spend $124.61 in electricity, as well as not being able to use your device for something else, plus the noise and heat from your device.

For $124, on Moonshot's official Kimi K3 API rates ($0.30 per 1M cached input, $3 per 1M fresh input, $15 per 1M fresh output), you can purchase 42 million fresh-input tokens, or 8.3 million generated output tokens, in whatever mix you want.

So what you get is 80x more expensive and you wait 416 days to get it.

Wow, you could do a lot at 292 tokens a sec—oh.

I have all praise for those taking this on and in my idiom would call it *the lord's work."

The image I reliably summon to mind is that compilation video showing the progress of Boston Dynamics bots. The curve between technically functional, to comically slow, to too slow for "real" work, on to, OMFG, may prove a (rough) curve.

It's work like this that moves things forward.

Hah I was looking for it and couldn't work out how many years/token. 292s is pretty good.

Seeing a lot of these “run 1TB models with 1GB RAM” projects recently. Most seem vibe coded and probably won’t be maintained.

Hoping a winner emerges with some real momentum behind it.

I love how goal posts are shifting from "vibe coded apps dont really work" to "vide coded projects wont be maintained"

You don't really need a maintainer when codex or claude code can set it up for you; thats how I got Trellis2 working on windows and tiny VRAM despite Microsoft recommending you have 24GB VRAM and Linux. Models are pretty disposable now.

So, instead of having one or two maintained projects that work super well, now we have hundreds or thousands of people half-assing it locally each time they need it? Doesn't sound efficient.

It might make sense to do this if there literally isn't an existing solution out there, but I think most would find that it's better to use an existing solution than to recreate the wheel every time there's a need. It's just less effort/time/money.

I have the same objection to forking open source projects; changes should be upstreamed when possible.

This is complete dependency on LLM tools all the way from development to usage, and is risky as well as prone to failure

Local inference without access to internet and tools is pretty safe, worst you get is a slop output, best case it solves your prompt.

No no, don't just say "vibe coded", say "Fable and $500 of credits"

Or Deepseek with a fiver.

Opus 4.6 was already enough to tackle these projects vibe coding.

Incidentally I think I was more productive with Opus 4.6 than with any subsequent Anthropic model.

Yea, opus 4.6 is much better than opus 5 because it is willing to do task. Opus 5 just gaslights and refuses to solve many complex problem.

these things can be done with local models; so don't say 500$ credits, say "accelerated climate change disasters by $500/year"

"This one is made with premium vibes"

How a patchwork of Python modules constitutes an API(?) or whatever(?) and how to use it (does anyone?), is beyond me. There is nothing that I would call documentation, let alone concrete usage examples. I know more amusing ways to waste my time. When I want to play with models, I use llama.cpp.

I love it how the rampocalypse is pushing people to squash all the performance they can. I hope that this also leads to rethinking model architecture so that they can do the same with less.

I’m still slightly confused on what this adds.

Let’s say I wanted to run a full size open weight model. I have a 128GB m3 max laptop.

Does this basically load layers in and out on demand? So I still have to download the full model to disk, but the RAM requirements go way down? The readme calls out that one still needs to connect HuggingFace, which leads me to believe that maybe you don’t even need to download the full model?

If you point it at a huggingface model identifier, it will download it, I assume. No way around that.

It reads like it is keeping only the core and the active layer loaded at any one point, and streams layers from disk; there are several other solutions like this and if my understanding is right, this is probably better than an mmap implementation or just streaming experts in.

Not an expert in this field, but the "expert" is consisting of multiple layers. To keep it small in terms of memory print, this project streams each layer (dividing even further).

It also requires extra space because of decomposition of the layers. Normally the file format optimized for compute intense workloads. But here the bottleneck is the memory capacity.

Also guessing that you need to be able to hold at least 3-layers at once in the memory, given M x N = R operation, M is the previous layer, N is next, and R is the result. on the next "layer", the R (result) becomes M, gets computed against the next layer, N, yielding the further result R'. And so on, until all layers are processed.

I assume it's horribly slow, but can be put in a non-intrusive background task...

Thanks for the extra detail here.

Obviously needs downloading eventually :).

It seems like this tool saves on both disk space and RAM, then. Classic trade off: speed vs space.

I mean, if you had fast enough internet connection, you could just stream it over https

You can run any frontier model on your PC if you just wait long enough...

What device do I need and how much will it cost to install one at home so that it works as quickly as the Claude Code answer (and it answers quite slowly)?

I guess the use case is something like: you have a slightly obsolete Mac or PC or a whole bunch of them, and just need to compose one or more convincing spam emails, but it's fine if it takes a full week to do it?

And you don’t pay for the electricity.

We will see if this project has legs. This is the kind of efficiency we desperately need. Now if we can address efficiency with llm training.

Running 70B on a 4GB GPU is wild. Really impressive engineering feat for resource-constrained environments.