I don't quite understand how people use these kind of local models that are specifically trained for agentic use cases. Like, what do they actually do with them? Do you use them as an assistant like with OpenClaw?
When used for coding, how do you divide the work between agents? Use a relatively fast model, like this one, as the main agent that calls out to other subagents that write the actual code? But then you'd need multiple models loaded, and why not just use the coding model to do everything? Do you use hosted cloud models for coding and code review, and this local one for orchestration? I don't see the benefit of that workflow either.
I've used local models to write some scripts etc, I like the idea of local computation instead of buying tokens from the cloud, but I don't know what the various agentic models are for, if not code generation.
I think the use case for this sort of model would be building something like Google or Apple's agents on device. Find an email for me, write a message for me, add an appointment to my calendar etc. Simple agentic stuff, not writing apps.
> it runs up to 1.7× faster than comparable MoEs on-device, with 184 tokens/s on an Apple M5 Max and up to 21K output tokens/s on a single NVIDIA B300 GPU
Wow, those are seriously impressive numbers for how intelligent this model seems to be!
One thing that this article doesn't clarify, and that I guess I really should look into, is what exactly AutomationBench tests for as compared to other benchmarks. The standout performance on AutomationBench is interesting to see, but I don't know what it really means because the model seems to score relatively close to other local models on every other benchmark.
I'd guess it's mostly just an artifact of the strength of their Agentic RL environments for those specific domains. If they had the same kind of access to massive amounts of institutional software knowledge, compute infrastructure, and hundreds of millions of users dropping them endless highly technical agent traces, there may have been similar disparities on the agentic benchmarks in other fields.
It's possible that they may have generally capped out for some reason... but, I'd rather be bullish on a new American AI lab making it to first place among western open-weight models that can run on a common consumer device. (Closest in size that's better across that range of benchmarks is Inkling-Small I think, at 276B params, unless I'm forgetting someone.)
Inkling last I used it was pretty terrible. I've never paid for a model that got stuck in a token puke loop and couldn't do tool calls until that day. Even the dated qwen models from almost a year ago 1/10th the size are a better user experience than the inkling model I piloted. Hoping they improve, sincerely, but not a good experience at all.
What does it matter what country an open weight model comes from? Please don't say you tried asking about tienamen square to a model 2 years ago and this was so scary you couldn't use it like 85% of people with this take have said...
You spin up your amazing startup in a day, full product and site.
It's got all the users embedded in a js file, but who cares.
Until people find out.
Then you task the agent with fixing it, it deletes prod.
You start over again.
It's beautiful, if Disney make movie about this I want to be in the credits.
Very cool. I’ve been on Gemma 4 12B because of my limited RAM. There’s already an Unsloth quantization of this model so I’m downloading that and will check its performance in the morning. Someone else in this thread recommended Ternary-Bonsai-27B, which I’ll also try.
Just also wanted to note that this model has a questionably-enforceable non-commercial use clause in its license. I really wish they had consulted with an IP lawyer, it’s very confusing.
Don't waste time on bonsai unless you have like 16gb or less ram (ram, not vram). Although I'm pretty sure you could still run this model well on 16+8, and qwen3.6 35b-a3b runs well on ~32gb combined ram at q4. Just offload most of the experts to cpu/ram if you have limited vram budget. With the moe models it doesn't hurt performance too much.
But bonsai is basically trash on real agentic coding work. I've tried both bonsai models, all the quants, messed with the hyperparameters, tried to fine tune (didn't succeed), and it just kept looping or failing to finish tasks. Qwen3.6 35b-a3b is my current go to for agentic work for any systems under 24gb vram, although honestly I'm not sure the dense qwen3.8 27b is meaningfully better. The main difference I see is it does a lot more testing attempts and burns a ton of tokens + compactions, so every task it takes 1.2-2x as long as 35b-a3b does. But the quality of the results is comparable.
I love your detailed explanation of things that I had little to no knowledge of, but I’m also stuck wondering if I’m using my setup correctly, as I have 16 GB of VRAM available but also 64 GB of regular RAM in my computer. Every time I use some setup or some configuration that overlaps or overflows between the two RAMs, my token per second drops significantly, me thinking that it is not set up correctly, so I’m stuck with 9 billion parameter models myself, even though I was under the impression with 64 GB of memory that I would be able to run something like a 30 billion model.
I do run the gpt-oss-20B fast on my 16GB vram, but it’s, limited.
Oh you should be running flash next with strata. Just go to stratas github page [0] and it will explain everything. It runs quite fast on 16+64, you should be able to get 50 tok/s or so. It's a 180b model and it feels a lot smarter than most of the other open weight options. As far as knowing "what will fit" in general that's hard to know these days. The moe models are hard to predict and run wildly different in different engines (llama.cpp, vllm, strata likes, etc) but they tend to be the only good options around 16gb vram. Dense models like qwen3.8 27b are only good if they're entirely in vram. Moe models you need the active count + a few experts (although it will run faster the more experts you can fit).
Thank you, very interesting. I also often use qwen 35b-a3b, when doing things locally, as I only have ~4.5GB free VRAM. My other goto is gemma 4 26b-a4b, which has faster prompt processing and the same token generation speed.
I've experimented with kolibri1 (78B-a3.5b), but despite the similar size in active parameters it is much slower. Even though the experts fit completely in main memory, as for the other models. I haven't tuned llama.cpp for it as much, so maybe there is still something to gain?
In a private benchmark I have, it performs much, much worse than Qwen 35B A3B for general knowledge. Still, probably beats Qwen 9B at much higher speed.
I asked it to implement the simplest todo in ghostty codebase and it failed badly - kept going in circles fixing one compile error and introducing another. Qwen 3.8 was slower (40 tok/s compared to 75) but did get it done.
in my experience it's stronger for terminal use, like chaining bash calls, monitoring background tasks, etc. rather than for straight up coding/swe work. Good with web search and file search rather than just editing software.
If Qwen had released a Qwen3.8 MoE of similar size, with a naive extrapolation its capabilities would roughly be similar to this model. So basically someone has caught up to Qwen (albeit Qwen from 1-2 months ago) in parameter efficiency.
I don't agree with that conclusion. I'm looking at the benchmarks page for LFM2.5-24B-A2B, and it appears to get a worse score on SWE-Bench-Verified (by a lot) when compared to Qwen 3.6 35B-A3B. It only clearly beats Qwen 3.6 35B on AutomationBench, ToolSandbox, and LiveClaw.
At least for my use cases, Qwen 3.6 35B-A3B appears better than this new release. An MoE for Qwen 3.8 would be even better than that!
Would you call Qwen 3.6 35B "similar size" with ~50% more parameters (both active and total) though? I believe what GP meant that if there was a Qwen 3.8 24B-A2B, it would roughly be as intelligent as this one.
yeah it's kind of the point. they're pedagogical tools. they give you the full recipe for building it from scratch or fine-tuning it beyond what it is.
but i'm prepared to be surprised, maybe they'll conjure up something standalone useful this time
I think their entire pitch about local AI is best demonstrated with their own model that they can optimise for it, and the shape of this model seems optimised for the smaller RTX chips. Indeed, this model appears to be part of their launch picture:
Maybe someone knows what order of cost a release like this requires in real terms of hardware and energy spend?
The cost over the diff from the moment of beginning training to the end. It's iterative from a prior model so it starts there and becomes this release after spending occurs I presume.
For example, is liquid spending millions in energy in this case and the cost is amortized over successive training runs over the initial datacenter spend? What if it's just eating cloud costs and has no dedicated hardware, that sort of thing. I don't particularly mean this model and liquid specifically. Just for this class of smaller fast on-device range of model what is the order of the training cost in real terms?
Maybe its short sighted to ask it as if it's a singular training run and is perhaps the end state of many variations of testing thus all of the R&D training runs would have to be included to to inform what became the final release. Seems like this should be taken into consideration. If anyone happens to know it'd be appreciated!
Training a mid-size model like this can take anywhere between $500k to like $5m. Liquid is spending money on GPU-hours (not sure if you'd call that "energy"). The amortization question is harder to answer without knowing their financials.
The license doesn't really make much sense to me, because if you're already at $10m revenue, you can probably train your own (better, more tailor-suited) model, or at the very least raise money to do so.
200-250 is the rate I've seen my context window fill up during longer tasks with Claude. The 40-80 figure is what Google's AI mode says, but I don't know where it got that from.
I've been messing around low computing power hardware to run models locally and to me 100 is already plenty fast. You most likely won't notice a huge difference between 184 and 250, unless you're running some serious multi-age agent circus.
The context window fills up also/mostly from prefil which is supposed to be 20+ times faster or so. I think 40-80 tps for decode/output is believable for claude models
Another data point: I got 94.74 tokens/s on an M4 Pro Mac mini 64GB with the simonw prompt "Generate an SVG of a pelican riding a bicycle". TTFT was 0.37s. The model thought for 40.75s before answering.
>How many tokens/s do you need for a coding agent to feel like Claude or codex?
it's kind of a loaded question, this model gets a lot of tok/s but reasoning is forced on; the time-to-answer is really a combination of things like "how much does this thing want to over-reason" and "how much does this thing like tool use?" .
You are right, they are not exactly the same thing, just wanted to have a reference point
When running locally usually I have to trade off speed for a better model. I can run decently big models at about 2tk/s, or a pretty dumb one at 40-100tk/s
AFAIU prefill speed makes a bigger difference in agentic coding workflows as the harness likely spends more time reading than creating new code. Qwen in particular spends a lot of time reasoning so it does make a larger impact there
When used for coding, how do you divide the work between agents? Use a relatively fast model, like this one, as the main agent that calls out to other subagents that write the actual code? But then you'd need multiple models loaded, and why not just use the coding model to do everything? Do you use hosted cloud models for coding and code review, and this local one for orchestration? I don't see the benefit of that workflow either.
I've used local models to write some scripts etc, I like the idea of local computation instead of buying tokens from the cloud, but I don't know what the various agentic models are for, if not code generation.
Wow, those are seriously impressive numbers for how intelligent this model seems to be!
One thing that this article doesn't clarify, and that I guess I really should look into, is what exactly AutomationBench tests for as compared to other benchmarks. The standout performance on AutomationBench is interesting to see, but I don't know what it really means because the model seems to score relatively close to other local models on every other benchmark.
It's possible that they may have generally capped out for some reason... but, I'd rather be bullish on a new American AI lab making it to first place among western open-weight models that can run on a common consumer device. (Closest in size that's better across that range of benchmarks is Inkling-Small I think, at 276B params, unless I'm forgetting someone.)
What does it matter what country an open weight model comes from? Please don't say you tried asking about tienamen square to a model 2 years ago and this was so scary you couldn't use it like 85% of people with this take have said...
Edit: I was wrong, I looked at the wrong model
Which makes for a dangerous combination, because it will delete prod, and it will do so at 250t/s.
You spin up your amazing startup in a day, full product and site. It's got all the users embedded in a js file, but who cares. Until people find out. Then you task the agent with fixing it, it deletes prod. You start over again.
It's beautiful, if Disney make movie about this I want to be in the credits.
Just also wanted to note that this model has a questionably-enforceable non-commercial use clause in its license. I really wish they had consulted with an IP lawyer, it’s very confusing.
But bonsai is basically trash on real agentic coding work. I've tried both bonsai models, all the quants, messed with the hyperparameters, tried to fine tune (didn't succeed), and it just kept looping or failing to finish tasks. Qwen3.6 35b-a3b is my current go to for agentic work for any systems under 24gb vram, although honestly I'm not sure the dense qwen3.8 27b is meaningfully better. The main difference I see is it does a lot more testing attempts and burns a ton of tokens + compactions, so every task it takes 1.2-2x as long as 35b-a3b does. But the quality of the results is comparable.
I do run the gpt-oss-20B fast on my 16GB vram, but it’s, limited.
[0] https://github.com/Niko1221/Strata
I've experimented with kolibri1 (78B-a3.5b), but despite the similar size in active parameters it is much slower. Even though the experts fit completely in main memory, as for the other models. I haven't tuned llama.cpp for it as much, so maybe there is still something to gain?
> What are signals in Angular?
> How does that relate to RxJS?
This was enough to loop Bonsai 2.
Same thing with Qwen3.6 35B A3B produced mangled garbage, but still, it didn't have errors in the code.
Exact same prompt in ChatGPT produces a half-decent model (ok, but exactly not the thing I asked) so long as animations are not involved.
fast is fast (18tk/s on a Intel Core 7 155H laptop), but quality is something else
Is that Qwen 3.8 27B? Because yes, you'd expect a plain 27B model to be better than a 24B-A2B MoE model all else being equal.
(All else isn't equal though of course - Qwen models are very, very good at coding)
Gemma-4 26BA4B would be a more similar comparison.
If Qwen had released a Qwen3.8 MoE of similar size, with a naive extrapolation its capabilities would roughly be similar to this model. So basically someone has caught up to Qwen (albeit Qwen from 1-2 months ago) in parameter efficiency.
At least for my use cases, Qwen 3.6 35B-A3B appears better than this new release. An MoE for Qwen 3.8 would be even better than that!
but i'm prepared to be surprised, maybe they'll conjure up something standalone useful this time
https://blogs.nvidia.com/blog/local-ai-rtx-spark-laptops-sof...
The cost over the diff from the moment of beginning training to the end. It's iterative from a prior model so it starts there and becomes this release after spending occurs I presume.
For example, is liquid spending millions in energy in this case and the cost is amortized over successive training runs over the initial datacenter spend? What if it's just eating cloud costs and has no dedicated hardware, that sort of thing. I don't particularly mean this model and liquid specifically. Just for this class of smaller fast on-device range of model what is the order of the training cost in real terms?
Maybe its short sighted to ask it as if it's a singular training run and is perhaps the end state of many variations of testing thus all of the R&D training runs would have to be included to to inform what became the final release. Seems like this should be taken into consideration. If anyone happens to know it'd be appreciated!
The license doesn't really make much sense to me, because if you're already at $10m revenue, you can probably train your own (better, more tailor-suited) model, or at the very least raise money to do so.
very nice.
I can't speak to the intelligence of it yet.
these small agentic-focused models are a lot of fun to play with if you haven't yet.
I get ~45 tokens/s generation and ~680 tokens/s prompt processing.
The problem is ollama is not caching the prompt, I get -> forcing full prompt re-processing due to lack of cache data
So this wouldn't work in its current state in an agentic loop.
That seems pretty decent. How many tokens/s do you need for a coding agent to feel like Claude or codex?
I've been messing around low computing power hardware to run models locally and to me 100 is already plenty fast. You most likely won't notice a huge difference between 184 and 250, unless you're running some serious multi-age agent circus.
(The resulting pelican was underwhelming.)
it's kind of a loaded question, this model gets a lot of tok/s but reasoning is forced on; the time-to-answer is really a combination of things like "how much does this thing want to over-reason" and "how much does this thing like tool use?" .
When running locally usually I have to trade off speed for a better model. I can run decently big models at about 2tk/s, or a pretty dumb one at 40-100tk/s
the results i've gotten so far have been a little braindead though. still playing around. bad luck with traversing filesystems remotely so far.
it seems to want to make concluding decisions very fast regardless of reasoning levels.