LFM2.5-8B-A1B
Liquid AI · 2026-05-28 · 8.5B parameters
LFM2.5-8B-A1B is Liquid AI's updated sparse MoE model with 8.3 billion total parameters and approximately 1.5 billion active per token, released in May 2026 as the next-generation successor to LFM2-8B-A1B. It is reasoning-tuned and emits an explicit chain of thought before the final answer. It retains the LFM2 hybrid architecture (18 gated LIV convolution blocks plus 6 GQA layers, 32 experts, top-4 routing) while extending training to 38 trillion tokens, increasing context to 128K tokens, and supporting 10 languages and Pythonic tool-call output. Reported at about 30 tokens per second on a phone and under 6 GB in laptop-class M5 Max and Ryzen tests, it is designed as an on-device personal-assistant model.