Liquid AI ships an 8B open MoE built for laptops and phones

Liquid AI released LFM2.5-8B-A1B, an open-weight mixture-of-experts model the company built for on-device agents. The "A1B" means 1 billion active parameters at inference time out of an 8 billion total, which is what lets it run on consumer laptops and phones at usable speed while keeping the capacity of a larger model. Pretraining scaled from 12 trillion tokens in the previous generation to 38 trillion, with RL phases targeting instruction following, hallucination reduction, and what the team calls "reasoning loop prevention."

The benchmark numbers Liquid points at are concentrated on agent-relevant capabilities rather than chat polish. The model hits 91.84 on IFEval, 88.76 on MATH500, and 88.07 on Tau² Telecom, an agentic benchmark where it edges past several larger closed models. Architecture-wise it combines grouped-query attention with gated short convolutions, and the context window has been pushed to 128K with the vocabulary doubled to 128,000 tokens for better multilingual coverage.

This is the part of the open model market that Apple, Microsoft, and Alibaba have all been pursuing for a year. Liquid's pitch is that an MoE that only loads a fraction of its weights per token is the cleanest way to fit useful tool calling into the memory and battery budget of a phone, and the 38T pretraining run signals they are willing to spend real money to make that pitch credible.

Why it matters

If you build local agents or assistants, this is now one of the strongest open candidates for the laptop and phone tier. The benchmark numbers point at tool calling specifically, which is the use case that makes or breaks an on-device agent. Test it on your real tool schemas before assuming the leaderboard transfers to your workload.