Model Overview & Capabilities
LFM2.5-24B-A2B is Liquid AI’s revolutionary on-device reasoning Mixture-of-Experts model released on October 13, 2026, securing Rank #28 on our global leaderboard with an index score of 28.8. Activating just 2.45 billion parameters per token from its 24-billion parameter base, it achieves verified scores of 29.0% on SWE-bench Verified, 28.5% on Terminal-Bench 2.0, 29.5% on Aider Polyglot, and 66.0% on GPQA Diamond.
The model is freely downloadable for private on-device execution on consumer hardware, with cloud endpoints available at $0.05 per million input tokens and $0.70 per million output tokens ($0.21 blended rate under 3:1 weighting). It delivers an astonishing inference velocity of 220 tokens per second with an ultra-responsive 0.48-second initial latency.
With a 131,072-token context window, LFM2.5-24B is purpose-built for local-first developer workstations, offline coding assistants, embedded systems, automotive edge computing, and privacy-sensitive corporate environments where proprietary source code cannot leave the local developer workstation. Its compact memory footprint allows software engineers to execute sophisticated reasoning loops entirely offline without any network connectivity.
