LFM2.5-230M
Liquid AI · 2026-06-25 · 230M parameters
LFM2.5-230M is Liquid AI's most compact LFM2.5 model, released June 25, 2026 for data extraction and lightweight on-device agentic pipelines rather than reasoning-heavy work. Its 14 layers combine eight gated short-convolution blocks with six grouped-query attention blocks at hidden size 1,024; the card gives a 32,768-token context, ten languages and a 19T-token training budget, while its base model's card says 28T. The BF16 checkpoint contains 229,693,184 parameters, and the tokenizer assigns 64,402 distinct IDs within 65,536 embedding rows.