Liquid AI Releases LFM2.5-24B-A2B, an Open Model for Local Agents

Liquid AI, an AI startup, has released LFM2.5-24B-A2B, a model with open weights and 24 billion parameters. It uses a mixture-of-experts design with 64 experts. Each token only passes through the top 4 of them. That means only 2 billion parameters activate per token. The model works with a context window of 131,000 tokens and was trained on 33 trillion tokens. On the AutomationBench agentic test, it scored 27.69 points. The Chinese model Qwen3.6-35B-A3B scored 8.72 on the same test. On an Apple M5 Max laptop, it runs at 184 tokens per second.

For simple, repetitive tasks, like filling out a form or calling the same tool over and over, a model like this on a laptop or phone actually makes sense. Privacy is a real plus here too, since the data never leaves the device. For serious agentic work, though, a model with only 2 billion active parameters is not in the same league as the larger models that lead the field today. And beating Qwen3.6-35B-A3B by almost three times on this one benchmark says more about that Chinese model's weakness than about this one's strength.