According to CoinMeta, Liquid AI has released the lfm2.5-24b-a2b model, which focuses on enhancing the local Agent's reasoning and tool invocation capabilities. The new version retains a total of 24 billion parameters, with only about 2 billion parameters being activated at a time. The amount of pre-training data has increased from 17 trillion to 33 trillion token, and the context size has been expanded from 32k to 131k. The new model can generate 184 token per second on Apple's M5 Max, with a speed that is up to 1.7 times that of similar MOE models. In Agent tests, liveclaw achieved a score of 48.09, which is higher than Qwen3.6-35b-a3b's score of 38.19 and more than three times that of Qwen's 27.69 on automationbench. However, in programming tests, its score was only 51%.