Liquid AI upgrades its 24B local model: activates only 2B parameters, outperforms larger Qwen in Agent tests.

动察 Beating AI News Flash: Frontier on-device AI R&D company Liquid AI has released LFM2.5-24B-A2B, with a focus on enhancing local Agent reasoning and tool-calling capabilities. The previous generation LFM2-24B-A2B mainly emphasized runtime efficiency. The new version retains the design of 24 billion total parameters with only about 2 billion parameters activated per inference, but pretraining data has increased from 17 trillion to 33 trillion tokens, and context has expanded from 32K to 131K.

The new model reasons before answering and is directly reinforcement-learned within Agent frameworks such as OpenClaw, Hermes Agent, Pi, and OpenCode, practicing continuous tool calling, handling returned results, and completing multi-step tasks. In an official demo, a single MacBook, without calling cloud model APIs, was able to have Hermes Agent search trending Hugging Face projects, compile a briefing, and continue operating based on follow-up questions.

Official tests show that the new model can generate 184 tokens per second on an Apple M5 Max, reaching up to 1.7 times the speed of comparable MoE models. In Agent tests, it scored 48.09 on LiveClaw, higher than Qwen3.6-35B-A3B's 38.19; on AutomationBench it scored 27.69, more than 3 times that of Qwen. However, on the SWE-bench Verified coding test, it scored only 51%, lower than Qwen's 66.2%.