What happened
On October 13, 2026, LiquidAI unveiled LFM2.5-24B-A2B, a reasoning model with 24 billion parameters, of which 2 billion are active, designed for quick, privacy-preserving agents running directly on devices. Both it and its LFM2.5-24B-A2B-Base variant have open weights and were posted to Hugging Face that same day. At launch, they worked with llama.cpp, MLX, vLLM, SGLang, and ONNX.
Why it matters
Developers can download both open-weight releases from Hugging Face and run them through popular inference tools from the start. The company says the model only performs reasoning, producing a visible step-by-step thought process before giving its answer, and supports a 131K-token context window.
Read the original at Hugging Face