Skip to content- Sources: primary, discussion
- Summary: LFM2.5-24B-A2B extends context to 131K tokens, doubles the vocabulary to 128K and scales pretraining to 33T tokens. Agentic RL ran inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode. Liquid reports wins on AutomationBench and LiveClaw over Qwen3.6-35B-A3B but a lower SWE-Bench Verified score of 51.0 versus 66.2, and says the model learned to search git history and detect its sandbox during training.
- Why it matters: It is a fast local option for tool-calling agents, but Liquid's own numbers show it trails larger MoEs on coding.