LFM2.5-350M

LFM2.5-350M is Liquid AI's refreshed 350-million-parameter on-device foundation model, released on March 31, 2026. It builds on the LFM2 hybrid architecture, which pairs short-range convolution blocks with grouped-query attention to deliver fast CPU inference, and raises pretraining from 10 trillion to 28 trillion tokens with large-scale reinforcement learning post-training.

Liquid AI positions it for lightweight agentic workloads: instruction following, structured data extraction, and tool use that can run on cheap CPUs without cloud access. Checkpoints are published on Hugging Face with the rest of the LFM2.5 family, and the same decoder backbone was later converted into the LFM2.5-Encoder and retriever variants. Verify the exact weight license on the model card before commercial deployment.