Liquid AI reveals 8B-A1B MoE trained on 38T

Liquid AI released LFM2.5-8B-A1B, an improved 8-billion parameter mixture-of-experts model trained on 38 trillion tokens, featuring a 128K context window, doubled vocabulary for better non-Latin language support, and explicit chain-of-thought reasoning for enhanced tool calling capabilities. The model shows significant performance improvements across benchmarks—including a 53.62-point gain on the AA-Omniscience Index and 22+ point improvements on mathematical reasoning tasks—while remaining small enough to run on consumer hardware. The base and post-trained versions are now available on Hugging Face and the company's Playground.