LFM2.5-24B-A2B beats Qwen3.6 at tools and loses SWE-Bench by 15 points

Twenty-four billion parameters, about two billion of them awake per token, and a laptop number on the box: 184 tokens a second on an Apple M5 Max. That's the pitch for LFM2.5-24B-A2B, which Liquid AI put on Hugging Face on 13 October. We read the full benchmark table and the license rather than the headline chart. It's a real tool-calling model. It's also a weak coding agent, and it isn't free for every company.

What the full table says

Liquid's launch chart leads with the wins, and some of them are big. AutomationBench at 27.69 against Qwen3.6-35B-A3B's 8.72 isn't a rounding error. LiveClaw goes 48.09 to 38.19, IFBench 74.69 to 64.21, IFEval 96.75 to 93.06. If your workload is "read this instruction, call the right function with the right arguments, don't wander off", these are the rows that matter, and they're good.

Now the rows further down. On BFCLv4, the function-calling benchmark most people check first, it scores 65.01 to Qwen's 65.00. That's a tie, not a win. PinchBench goes the other way (83.18 vs 89.93), and so does WebArena Verified (26.75 vs 37.62), which is the one closest to "drive a browser for me". Then SWE-Bench Verified: 51.00 against 66.20. Nemotron-3.5-Lightning, also in Liquid's own table, posts 53.40 there. So among the four models Liquid chose to compare against, its new one is third on repo-level coding.

Knowledge is the other soft spot, and Liquid says so itself. GPQA Diamond lands at 74.42 versus 83.69 for Qwen, and the blog recommends pairing the model with retrieval or web search for knowledge-heavy work. We'd take that advice literally. A 2B-active model that's been tuned hard for instruction following will answer confidently. It won't always know.

The laptop numbers, and what they leave out

The 184 tokens per second on an M5 Max and 114 on an AMD Ryzen AI Max+ 395 are the reason most people will try this. They're plausible for a model that only fires about 2B parameters per token. But Liquid's announcement doesn't say which quantization produced them, how long the prompt was, or how much memory the run took. That matters more than usual here, because all 24B parameters still have to sit in memory even though only a slice of them computes. The datacenter figure is better documented: 21,000 output tokens a second on one Nvidia B300, with 1,024 input tokens, up to 256 output tokens and BF16 weights.

There's a second catch that won't show up in tokens per second. It's reasoning-only. Liquid describes it as producing an explicit chain of thought before every final answer, and neither the blog nor the model card mentions a switch to turn that off. Fast tokens help less when the model spends a chunk of them thinking. We haven't timed it ourselves, so we can't tell you how big that chunk is on a typical tool call, and we'd want to see it before calling it "fast" for interactive use.

The architecture changes from LFM2-24B-A2B, the February release, are real: context goes from 32K to 131,072 tokens, the tokenizer from 65K to 128K entries, and pre-training ran to roughly 33 trillion tokens. Same 24B total, same 2B-active shape. If you already run LFM2 in llama.cpp, the new GGUF should slot in, but re-check your context settings, since the old 32K cap no longer applies.

License and price, before you ship it

The model card's license tag reads "lfm1.0", which is the LFM Open License v1.0. Its commercial grant is conditioned on not crossing a threshold, defined as annual revenue of $10,000,000 or more. Below that, you can use it commercially. Above it, you need a separate deal with Liquid. Qualified non-profits are exempt for non-commercial and research use. That's a lot more permissive than a research-only license, and a lot less than Apache 2.0. If you work at a company past that line, a self-hosted pilot is a legal question before it's a technical one.

If you'd rather not host it, Liquid's own console lists $0.10 per million input tokens, $0.05 for cached input and $1.00 per million output tokens, and OpenRouter shows the same rates. The 10:1 output ratio is steep for a 2B-active model, and since every answer comes with reasoning tokens, output is where the bill lands. We've added the model to our open-weight models tracker, and you can compare the hosted price with the bigger names on the LLM API pricing tracker.

Our take: for local agents that mostly call functions and follow strict instructions, it's worth a test this week, and it might replace a 30B-class model on a single workstation. For coding agents, it isn't the pick, by Liquid's own numbers. I might be wrong about the laptop experience until someone publishes the memory and quantization behind that 184.

FAQ

Sources: Liquid AI, LFM2.5-24B-A2B announcement (13 October 2026), model card, LFM Open License v1.0 and Hugging Face blog post; independent: OpenRouter model page and MarkTechPost. All read on 14 October 2026.