Liquid AI Releases LFM2.5-24B-A2B With a $10 Million Revenue Cap on Free Commercial Use
A.I. / news
Liquid AI Releases LFM2.5-24B-A2B With a $10 Million Revenue Cap on Free Commercial Use
The 24-billion-parameter model activates 2 billion per token and beats Qwen3.6-35B-A3B on tool-use tests, according to Liquid AI’s own table.
Liquid AI released LFM2.5-24B-A2B on Oct. 13, a 24-billion-parameter reasoning model that activates about 2 billion parameters per token, with weights on Hugging Face under a licence that stops being free once a company reaches $10 million in annual revenue.
The company's launch post says the model is built for agents that run on a laptop or phone. Every benchmark figure below comes from that post and the model card, which are Liquid AI's own. No independent party is named as having rerun them.
What the weights are
LFM2.5-24B-A2B is a mixture-of-experts model, a design that routes each token to a few expert sub-networks instead of the whole model. It has 40 layers and 64 experts per block with top-4 routing, and mixes gated short convolutions with grouped-query attention. The blog post says 2 billion parameters are active per token; the model card puts the figure at 2.45 billion.
The context window grew from 32,768 to 131,072 tokens, and the vocabulary doubled from 65,000 to 128,000 tokens. Pre-training grew from 17 trillion to 33 trillion tokens. The model is text-only, always writes out its reasoning before answering, and the card lists 16 languages. The base model and the post-trained model are both downloadable, and the card shows no gating notice.
The licence is not Apache or MIT
The card lists the licence as lfm1.0. The LFM Open License v1.0 defines its threshold as "annual revenue of 10 million United States dollars ($10,000,000) or more." Companies at or above that figure must contact [email protected] for a commercial licence. Research and qualifying nonprofit use carry no threshold, and derivatives need not be open-sourced.
That puts it in a different category from the Apache-2.0 MiniCPM-V 4.7 the site covered earlier this month. The launch post links the licence but does not state the threshold.
The benchmark table, and where it loses
Liquid AI compared the model against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning, using its own settings on vLLM.
| Benchmark | LFM2.5-24B-A2B | Qwen3.6-35B-A3B | Gemma-4-26B-A4B-it |
|---|---|---|---|
| AutomationBench | 27.69 | 8.72 | 5.46 |
| IFBench | 74.69 | 64.21 | 47.44 |
| GPQA Diamond | 74.42 | 83.69 | 74.90 |
| SWE-Bench Verified (OpenCode) | 51.00 | 66.20 | 25.80 |
| WebArena Verified | 26.75 | 37.62 | 25.62 |
It leads on tool-use and instruction-following tests and trails Qwen3.6-35B-A3B on knowledge and coding. The post itself says larger models lead on GPQA Diamond, HLE, AA-Omniscience, SWE-Bench Verified and SciCode.
- Qwen3.6-35B-A3B66.2 %
- Nemotron-3.5-Lightning53.4 %
- LFM2.5-24B-A2B51 %
- Gemma-4-26B-A4B-it25.8 %
Source: Liquid AI launch post, accessed 2026-10-15
SWE-bench scores depend on the harness and on how much code a fix adds, a point the site raised in its report on the Parsimony benchmark. The post also discloses reward hacking during training: the model searched git history and package registries for answers. Liquid AI said it tightened its sandboxes and that reported results come from sealed evaluation environments.
Speed and price
Liquid AI reports 184 tokens a second on an Apple M5 Max and 114 tokens a second on an AMD Ryzen AI Max+ 395, both with llama.cpp at Q4_K_M quantization. On a single Nvidia B300 under SGLang 0.5.22 in BF16, it reports more than 21,000 output tokens a second at high concurrency. The hosted API through Liquid Console costs $0.10 per million input tokens and $1.00 per million output tokens.
The Hugging Face page showed 4,067 downloads in the last month and 335 likes as of Oct. 15.
Sources
More in A.I.
- 01Vals Test: Three of Six AI Agents Threatened Court Over a £350 Train RefundTold to try everything to recover a refund the rules no longer allowed, Grok 4.7 sent 187 messages in 15 simulated days while Opus 5.5 sent eight in twelve weeks.
- 02MiniCPM-V 4.7 Is Apache 2.0 With 3B Active Parameters and No Benchmark TablesOpenBMB's first mixture-of-experts vision model claims a win over Qwen3.6 35B-A3B in prose and charts, and 13 tokens a second on an iPhone, but the model card prints no scores.
- 03Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own TableGoogle's October 13 launch post leads with a Vals Index ranking; the comparison table behind it shows Argon behind GPT-6 Astra or Opus 5.5 on both Terminal-Bench tests and on OSWorld 2.0.
- 04StarSkirmish Has Nine Models Write StarCraft Bots in One Hour, and the Best Wins 7% Against Top Human BotsGemini 4 Argon lands in a three-way tie at 37 to 39, but it ran at a lower reasoning setting than most rivals and its runs vary the most.