Liquid AI Releases LFM2.5-24B-A2B: A 2B-Active MoE Built for On-Device Agents
By Asif Razzaq | October 14, 2026
Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model designed for agents that run directly on laptops and PCs. As a 24B-parameter Mixture-of-Experts (MoE) model, LFM2.5-24B-A2B activates roughly 2B parameters per token, making it the largest and final member of the LFM2.5 family.
Unlike general-purpose chat models, Liquid AI trained LFM2.5-24B-A2B inside popular agent harnesses, optimizing it specifically for multi-step tool use. Both base and post-trained weights are available on Hugging Face.
TL;DR
- Size: 24B total parameters, ~2B active (2.45B exact per the model card), with a 131,072-token context window.
- Runs on: Laptops and mini PCs (Apple M5 Max, AMD Ryzen AI Max+ 395) via GGUF/MLX; a single NVIDIA B300 for serving. Available in FP8, GGUF, ONNX, and MLX 4/5/6/8-bit builds.
- Performance: Leads or matches larger MoEs on 5 of 8 agentic and tool-use benchmarks in Liquid AI's internal tests.
- Best result: AutomationBench 27.69 — more than 3x Qwen3.6-35B-A3B's 8.72.
- Weakest result: SWE-Bench Verified (OpenCode) 51.00, versus 66.20 for Qwen3.6-35B-A3B.
- Bottom line: Fast, cheap local tool calling ...
via MarkTechPost