Liquid AI released LFM2.5-24B-A2B on October 13, a 24-billion-parameter mixture-of-experts reasoning model that it pitches at agents running on laptops and edge hardware. Liquid AI’s announcement The model card lists 2.45 billion active parameters per token, which the blog rounds to 2B. Model card In Liquid’s own tables it beats Qwen3.6-35B-A3B and Gemma-4-26B-A4B-it on most tool-use and instruction-following tests while losing to Qwen on knowledge and coding. Liquid AI’s announcement Whether you can ship it depends on the license: the weights carry the LFM Open License v1.0, which stops covering commercial use once the company using it reaches $10 million in annual revenue. LICENSE file
S5 Labs has not run the model. Every benchmark and speed figure below is Liquid’s.
What Liquid shipped
The architecture is unchanged from LFM2-24B-A2B: 40 layers, 30 of them gated short-convolution blocks and 10 grouped-query attention, with 64 experts per MoE block and four routed per token. Liquid AI’s announcement Model card What changed is the training and the context length. The context window goes from 32K to 131,072 tokens, and the vocabulary doubles to 128,000 through an in-place tokenizer expansion. Model card
| Item | LFM2.5-24B-A2B |
|---|---|
| Parameters | 24.0B total, 2.45B active |
| Context | 131,072 tokens |
| Training | ~33T pre-training tokens (up from ~17T), then 2T mid-training at 32K and 300B at 131K |
| Languages | 16, including Arabic, Chinese, Hindi, Japanese, Korean, Thai and Vietnamese |
| Variants | Base, FP8, GGUF, ONNX, MLX 4-bit; SGLang draft model “coming soon” |
| Recommended sampling | temperature 0.1, top_k 50, repetition_penalty 1.05 |
Sources: model card, announcement.
Liquid says post-training included agentic reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode. Liquid AI’s announcement The model always reasons before answering; the chat template opens a <think> block, so budget for those tokens in latency and context. Model card Day-one runtimes are llama.cpp, MLX, vLLM, SGLang and ONNX, and the Transformers example needs transformers>=5.0.0. Liquid AI’s announcement Model card
Where it leads, and where it doesn’t
Liquid ran every model in its comparison on vLLM with what it calls recommended generation parameters, mostly temperature 0.6 and 32,768 maximum output tokens. Model card A selection of its rows:
| Benchmark (Liquid’s runs) | LFM2.5-24B-A2B | Qwen3.6-35B-A3B | Gemma-4-26B-A4B-it | Nemotron-3.5-Lightning |
|---|---|---|---|---|
| AutomationBench | 27.69 | 8.72 | 5.46 | 2.61 |
| LiveClaw | 48.09 | 38.19 | 41.25 | 37.72 |
| ToolSandbox | 80.71 | 75.02 | 72.09 | 75.13 |
| BFCLv4 | 65.01 | 65.00 | 57.09 | 53.78 |
| IFBench | 74.69 | 64.21 | 47.44 | 71.67 |
| IFEval | 96.75 | 93.06 | 92.00 | 94.19 |
| PinchBench | 83.18 | 89.93 | 77.12 | 84.20 |
| WebArena Verified | 26.75 | 37.62 | 25.62 | 28.69 |
| GPQA Diamond | 74.42 | 83.69 | 74.90 | 72.42 |
| SWE-Bench Verified (OpenCode) | 51.00 | 66.20 | 25.80 | 53.40 |
Source: Liquid AI’s announcement.
Liquid’s model is strongest at calling tools and following instructions, and its AutomationBench lead is wide: 27.69 against 8.72 for the next model. BFCLv4 is a tie at a 0.01-point margin. On browser tasks, coding and graduate-level science, Qwen3.6-35B-A3B leads, and Nemotron-3.5-Lightning also edges out LFM2.5 on SWE-Bench Verified, PinchBench and WebArena Verified. Liquid’s own blog says larger models lead on knowledge and coding and suggests pairing LFM2.5 with retrieval or web search. Liquid AI’s announcement
The speed figures come from narrow test settings
Liquid reports 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, and says that is up to 1.7 times faster than the next-best model it compared, using the least memory. Liquid AI’s announcement Those edge figures are decode speed from llama.cpp with Q4_K_M weights and a 4K-token input. Model card Agent loops routinely carry far longer contexts than 4K, so expect lower numbers once tool results and history pile up.
The server figure, over 21,000 output tokens per second on one NVIDIA B300, comes from SGLang 0.5.22 in BF16 with 1,024 input tokens, at most 256 output tokens per request, and a sustained pool of concurrent requests whose size the card does not state. Model card It is a throughput number for short jobs, not a measure of how fast one user sees a reply.
The license decides it for many teams
The LICENSE file in the repository defines a threshold of $10,000,000 or more in annual revenue. Above it, commercial use “is not licensed under this Agreement”. LICENSE file The text is loose at the boundary, saying “or more” in one place and “exceeds” in another, so treat $10 million itself as the paid side. Revenue counts across the whole legal entity, which the license defines to include anything that controls the user, is controlled by it, or sits under common control, with control meaning management power, 50% or more of shares, or beneficial ownership. The license does not say how revenue is measured. Rights end automatically on non-compliance. LICENSE file
Liquid’s license help page says the license is “based directly on Apache 2.0”, with the threshold as the main difference. It also says fine-tuned and modified models stay under the same license and threshold, and that the license does not automatically cover a downstream customer’s commercial use; customers over $10 million need their own commercial license, through [email protected]. Liquid AI, model license The revenue threshold does not apply to qualified non-profits (501(c)(3) or equivalent) using the models for non-commercial or research purposes; for-profit companies stay under it. Liquid AI, model license
The downstream-customer rule is the one product builders need to read. A startup under the line can ship LFM2.5 inside an app, but if that app goes to an enterprise customer, the customer’s revenue may put it on the paid side. Two of the models in Liquid’s comparison avoid this: Qwen3.6-35B-A3B and Gemma-4-26B-A4B-it are tagged Apache 2.0 on their model cards. Qwen3.6-35B-A3B Gemma-4-26B-A4B-it Liquid’s d1-3B decision model ships under the same LFM license, so the same question applies there.
What to test before choosing
- Run your own tool schemas through the model at the context length your agent reaches after ten or twenty turns, not at 4K.
- Count the thinking tokens. A model that always reasons can lose its decode-speed advantage on short tool calls.
- Compare against Qwen3.6-35B-A3B on the same harness. If your agent writes code or browses, Liquid’s own table favors Qwen.
- Check the memory you actually have. Liquid gives no file size or runtime memory figure for the quantized builds, and total parameters, not active ones, set the weight size. Our local LLM machines guide covers sizing.
- Ask legal or procurement whether your entity, its parent and your likely customers sit under $10 million before an evaluation becomes a dependency.
What is still unknown
No one outside Liquid has published reproductions of these benchmarks or speeds yet. Liquid has not given quantized file sizes or memory use, and the speculative-decoding draft model is not out. The license page names Liquid Foundation Models in general rather than this checkpoint; the coverage comes from the lfm1.0 tag and LICENSE file in the repository. For a team under the revenue line building a tool-heavy local agent, LFM2.5-24B-A2B is worth a test this week. For anyone selling to larger companies, the Apache 2.0 alternatives carry less paperwork, and the benchmark gap would need to show up in your own harness to justify it.