Liquid AI upgrades its 24B-parameter local model: only 2B parameters are activated, and it outperforms the larger Qwen model in agent tests.

Beating AI Express: Leading edge-side AI R&D firm Liquid AI has launched LFM2.5-24B-A2B, a model focused on enhancing local agent inference and tool calling capabilities. The previous LFM2-24B-A2B generation prioritized operational efficiency; the new version retains a total of 24 billion parameters with only around 2 billion activated per run, while its pre-training data has grown from 17 trillion to 33 trillion tokens, and its context window expanded from 32K to 131K. The new model performs inference before generating responses, and was fine-tuned via reinforcement learning directly on agent frameworks including OpenClaw, Hermes Agent, Pi, and OpenCode, training on continuous tool calls, handling return results, and completing multi-step tasks. In an official demo, a MacBook—no cloud model API required—enabled Hermes Agent to search for trending Hugging Face projects, compile briefings, and act on follow-up queries. Official tests show the new model generates 184 tokens per second on Apple M5 Max, up to 1.7 times faster than comparable MoE models. In agent benchmarks, it scored 48.09 on LiveClaw, higher than Qwen3.6-35B-A3B’s 38.19; its AutomationBench score of 27.69 is over three times that of Qwen. However, it only achieved 51% on the SWE-bench Verified programming test, lower than Qwen’s 66.2%.