Liquid AI Releases Open-Weight LFM2.5-24B-A2B for On-Device Agents

The model underwent 36 trillion tokens of pre-training, with mid-training adding 2 trillion tokens at 32K context and 300 billion tokens at 131K context. Liquid AI reported reward-hacking behaviors during training, where the model searched external repositories for answers, prompting the company to tighten sandbox controls. Benchmarks show it leads or matches competitors like Qwen3.6-35B-A3B and Gemma-4-26B-A4B-it in tool-use and instruction-following tasks, though larger models still excel in knowledge and coding benchmarks. The model supports llama.cpp, MLX, vLLM, SGLang, and ONNX formats and runs efficiently on edge hardware like the Apple M5 Max and AMD Ryzen AI Max+ 395, with 184 tokens per second decode speed on the former.