Hugging Face Releases Edge Inference Model LFM2.5-24B-A2B

On October 14, 2026, Hugging Face released the LFM2.5-24B-A2B model, an on-device reasoning model featuring explicit chain-of-thought capabilities and a 131K context window. Building upon the LFM2-24B-A2B architecture, the new model retains 40 layers, 64 experts per MoE block with top-4 routing, and the hybrid design of gated short convolutions and GQA. Its context window expands significantly from 32K to 131K tokens, while an in-place tokenizer expansion doubles the vocabulary from 65K to 128K tokens to improve non-Latin tokenization. Unlike its predecessor, LFM2.5-24B-A2B is a reasoning-only model that produces an explicit chain of thought prior to generating its final answer. Its pre-training scale was increased from approximately 17T to roughly 33T tokens. The mid-training phase incorporates an additional 2T tokens at a 32K context weighted toward reasoning, math, code, instruction following, and multilingual text, alongside 300B tokens at a 131K context focused on long-context data. Tool calls and multi-turn trajectories are integrated throughout both training phases, and the model is specifically designed for efficient on-device execution and agentic tool use.