LFM2.5-24B-A2B

LFM2.5-24B-A2B is the largest and final model in Liquid AI's LFM2.5 family, released October 13, 2026. It totals 24 billion parameters but activates only 2 billion per token across 40 layers and 64 experts per MoE block with top-4 routing. Both the instruct and base checkpoints are published as open weights on Hugging Face.

The model always reasons before answering: it writes out a chain of thought, then calls tools and handles multi-step tasks. Context grew to 131k tokens and vocabulary to 128k for better non-Latin language handling, with pretraining scaled to about 33 trillion tokens. Post-training used agentic reinforcement learning inside harnesses like OpenClaw, Hermes Agent, Pi, and OpenCode. Vendor benchmarks claim leadership on AutomationBench and LiveClaw versus Qwen3.6-35B-A3B, which are company-reported figures worth independent verification. Verify the exact weight license on the model card before shipping.