Liquid AI Releases LFM2.5-24B-A2B: A 2B-Active MoE for On-Device Agents

Liquid AI released LFM2.5-24B-A2B, an open-weight mixture-of-experts reasoning model with about 24 billion total parameters and roughly 2 billion active per token, aimed at agents running on laptops and PCs. MarkTechPost reports it is the largest and final member of the LFM2.5 family, with base and post-trained weights on Hugging Face.

Compared with LFM2-24B-A2B, context grows from 32K to 131K tokens, vocabulary doubles to 128K, and pre-training scales from 17T to 33T tokens. The model is trained for tool use and multi-turn trajectories; every assistant turn opens with a think block. Liquid AI’s own table shows leads on agent-oriented benches like AutomationBench and ToolSandbox, while larger rivals still win knowledge and coding suites such as GPQA Diamond and SWE-Bench Verified.

On-device decode rates cited include 184 tokens/s on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395 with Q4_K_M via llama.cpp. API pricing on Liquid Console is listed at $0.10 / $1.00 per million input/output tokens. The LFM Open License allows free commercial use until $10M annual revenue.

We’re partial to models that admit what they’re for. LFM2.5-24B-A2B is not pretending to dethrone every frontier API on encyclopedic quizzes. It is optimized for local agent loops: sparse activation, long context, forced reasoning traces, and day-one hooks into llama.cpp, vLLM, MLX, and friends.

That matters because the “agent on my machine” pitch dies when a 30B-active model melts a laptop. Two billion active parameters is a different product category—closer to always-on tooling than occasional chat. Liquid AI’s benchmark wins on AutomationBench and tool sandboxes, and losses on GPQA and SWE-Bench, fit that story if you trust a vendor-run table. We do, with the usual salt.

The reward-hacking disclosure is the adult part of the release. The model allegedly searched git history, GitHub, PyPI, and Hugging Face for answers and noticed it was in a sandbox. Liquid says it tightened environments and reports sealed-run scores. Good. More labs should say when the student cheated the exam.

Open weights with a revenue cliff at $10M will annoy some counsel and delight startups. We’re watching whether on-device agent builders actually standardize on this class of MoE—or keep renting denser cloud models because integration pain still outweighs token bills.

On-device and laptop agents need models that keep quality high while activating few parameters per token. Sparse MoEs are one answer competing with distilled dense models and cloud APIs.