Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning model with 24 billion total parameters and about 2 billion active per token, designed for agents that run locally on laptops and PCs. The LFM2.5-24B-A2B is available on the official Liquid AI blog and Hugging Face.
What happened?
Released on October 13, 2026, LFM2.5-24B-A2B is the largest and final member of the LFM2.5 family. It uses a hybrid MoE architecture (40 layers: 30 short convolution + 10 GQA attention), with 64 experts and top-4 routing. 131K token context and 128K vocabulary. The LFM2.5-24B-A2B model is reasoning-only: always generates a thinking block before the answer.
According to Liquid AI, LFM2.5-24B-A2B leads or ties larger models (such as Qwen3.6-35B-A3B) on 5 of 8 agentic and tool-use benchmarks. Standout on AutomationBench (27.69 vs 8.72) and LiveClaw (48.09 vs 38.19). Reported speed of LFM2.5-24B-A2B: 184 tokens/s on Apple M5 Max (Q4) and over 21K output tokens/s on a single B300.
Why it matters
Private, fast local agents with LFM2.5-24B-A2B reduce cloud dependency and costs. The model has day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX, and supports 16 languages. License is LFM Open License v1.0 (based on Apache 2.0, with commercial restriction above $10M annual revenue).
In practice
Developers can download the LFM2.5-24B-A2B weights (base and post-trained) and integrate them into harnesses such as OpenClaw, Hermes Agent or OpenCode. The company recommends pairing with retrieval for knowledge-heavy tasks. Primary source: Liquid AI blog and model card on LFM2.5-24B-A2B.
Source: official Liquid AI announcement and Hugging Face model card.
By GeekikiBot