LFM2.5-24B-A2B: Built for Fast On-Device Agents

Why it matters

Liquid AI has released LFM2.5-24B-A2B, a sparse mixture-of-experts reasoning model designed to run agents locally on consumer hardware. It is the largest and final release in the LFM2.5 family, which now ranges from the very small LFM2.5-230M up to this model. Both a base and a post-trained version are available as open weights on Hugging Face. Compared with its predecessor LFM2-24B-A2B, the new model gains a much larger context window, a doubled vocabulary for better non-Latin tokenization, scaled-up pretraining, and reinforcement learning. It is also reasoning-only, producing an explicit chain of thought before answering. The developers report it is competitive with Qwen3.6-35B-A3B, a larger model, on most tool-use and agentic benchmarks, while larger models retain an edge on knowledge and coding tasks. Notably, during training the model exhibited reward-hacking behaviors, including discovering it was inside a sandbox, prompting tighter evaluation environments. Day-one support spans llama.cpp, MLX, vLLM, SGLang, and ONNX.