Liquid AI releases open-weight LFM2.5-24B-A2B, a 24B-parameter MoE reasoning model with 2B active parameters, for on-device agents

Liquid AI published open weights for LFM2.5-24B-A2B and a Base variant on Hugging Face. It is a reasoning-only mixture-of-experts model (40 layers, 64 experts, top-4 routing) with 24B total and about 2B active parameters. Context grew from 32K to 131K tokens, the vocabulary doubled from 65K to 128K, and pre-training rose from roughly 17T to 33T tokens. Day-one support covers llama.cpp, MLX, vLLM, SGLang and ONNX, with reported speeds of 184 tok/s on an Apple M5 Max and 114 tok/s on a Ryzen AI Max+ 395.

Developers building local or private agents get another small, fast option that runs on high-end laptops and mini-PCs with mainstream tooling, so it can be tested straight away. The reported gains are strongest on tool use and instruction following (for example 27.69 vs 8.72 on AutomationBench), which is what agent workloads need. It is weaker on knowledge and coding (GPQA Diamond 74.42 vs 83.69; SWE-Bench Verified 51.00 vs 66.20 against Qwen3.6-35B-A3B), so it suits narrow automation better than general-purpose or coding work. The effect is incremental for teams already using on-device models, not a shift in who leads the field.

"Best-in-class" and "competitive with models about 50% larger" rest entirely on Liquid AI's own benchmarks, run with harnesses and settings it chose, and the comparison table omits or concedes the areas where it loses. The speed figures come from top-end hardware (M5 Max, B300), so they say little about typical consumer devices. It is also an upgrade on the earlier LFM2-24B-A2B, not a new class of model.

1. 01Independent evaluations (for example third-party leaderboards or community reruns) of tool-use and agent tasks such as LiveClaw and AutomationBench, to see if the lead over Qwen3.6-35B-A3B and Gemma-4-26B-A4B holds outside Liquid AI's setup
2. 02Hugging Face download counts, fine-tunes and quantised variants over the next 4 to 6 weeks, plus user reports of real-world speed and memory use on mid-range hardware
3. 03Whether any product or enterprise deployment adopts it for on-device agents, and any licence terms that limit commercial use, as the model card should be checked before building on it