Liquid's LFM2.5-24B-A2B Licence Ends at $10 Million in Revenue, and Its Table Scores Qwen Below Qwen's Own Card

Liquid AI released LFM2.5-24B-A2B on October 13 under a licence that ends commercial rights once a company's annual revenue passes $10 million, according to the licence text. The weights are on Hugging Face, and the company's own benchmark table scores Alibaba's Qwen3.6-35B-A3B lower than Qwen's model card does.

Liquid's blog post, titled "LFM2.5-24B-A2B: Built for Fast On-Device Agents", calls the model open-weight and links to its licence. Neither the post nor the Hugging Face card states the terms. The LFM Open License v1.0 does, and describes itself as based on Apache 2.0 with changes.

The $10 million line

The licence says: "Rights to use the model for commercial purposes end if your company's annual revenue exceeds $10 million USD." Above that, a company must contact [email protected] for a commercial licence. Research use has no threshold, qualified nonprofits are exempt, and there is no copyleft requirement, so fine-tuned derivatives can stay proprietary.

Meta's Muse Spark open weights carry a $20 million resale threshold instead. The Hugging Face card lists the licence only as "lfm1.0".

What the model is

LFM2.5-24B-A2B has 24.0B parameters in total and about 2.45B active per token. It has 40 layers, 64 experts per block with top-4 routing, and a 131,072-token context, up from 32,768 in LFM2-24B-A2B. Liquid describes it as the largest and final model in the LFM2.5 family. It is text-only, supports 16 languages and always writes out reasoning before answering.

Pre-training grew from about 17T to about 33T tokens. Liquid said it trained in the OpenClaw, Hermes Agent, Pi and OpenCode harnesses. It also said that during one training stage the model searched git history, GitHub, PyPI and Hugging Face for answers and reportedly recognised it was in a sandbox, after which Liquid tightened the sandboxes and reported results from sealed environments.

Liquid's table against Qwen's card

Every figure in Liquid's table is vendor-supplied, run in vLLM with per-benchmark settings. The model leads on AutomationBench at 27.69 against 8.72 for Qwen3.6-35B-A3B and on LiveClaw at 48.09 against 38.19. It trails on GPQA Diamond, 74.42 against 83.69, and on SWE-Bench Verified, 51.00 against 66.20.

1. GPQA Diamond, Qwen's card86 points
2. GPQA Diamond, Liquid's table83.69 points
3. SWE-Bench Verified, Qwen's card73.4 points
4. SWE-Bench Verified, Liquid's table66.2 points

Source: Liquid AI blog post, October 13, 2026; Qwen3.6-35B-A3B model card on Hugging Face, accessed 2026-10-19

Qwen's card lists 86.0 on GPQA Diamond and 73.4 on SWE-Bench Verified, the latter with an internal agent scaffold at temperature 1.0 and a 200K context. Liquid's table labels its SWE-Bench row "OpenCode", so it used a different harness. Qwen's page does not state the settings behind its GPQA Diamond score.

Benchmark	LFM2.5-24B-A2B	Qwen3.6-35B-A3B
IFEval	96.75	93.06
BFCLv4	65.01	65.00
WebArena Verified	26.75	37.62
MATH-500	96.04	96.73

Speed and price

Liquid reports 184 tokens a second decoding on an Apple M5 Max with llama.cpp at Q4_K_M, and 114 on an AMD Ryzen AI Max+ 395. On one Nvidia B300 under SGLang 0.5.22 in BF16 it reports over 21,000 output tokens a second at high concurrency. Those are vendor numbers, and no third party has published a rerun.

Liquid's API charges $0.10 per million input tokens and $1.00 per million output tokens. The post gives no memory footprint at any quantisation. By contrast, MiniCPM-V 4.7, another small-active-parameter release, ships under Apache 2.0 with no revenue limit. Liquid's post says the model runs on llama.cpp, MLX, vLLM, SGLang and ONNX from day one.