LFM2.5-24B-A2B: Built for fast on-device agents

The model has 24 billion parameters but activates about 2.45 billion per token. It uses 64 experts per MoE block with top-4 routing. Context grows from 32,768 to 131,072 tokens over its predecessor. In Liquid's own tests, it scored 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B, and 27.69 on AutomationBench against 8.72. Reported speeds are 114 tokens per second on AMD's Ryzen AI Max+ 395. Weights come in native, FP8, GGUF, ONNX and MLX-4bit variants under the lfm1.0 license. The gains are narrow. It trails Qwen3.6-35B-A3B on knowledge and coding benchmarks, so it suits agent plumbing better than heavy code generation. Liquid ran agentic reinforcement learning inside OpenClaw, Hermes Agent, Pi and OpenCode. It is the largest and final member of the LFM2.5 family. Note that 24B of weights still means far more memory than a 2B model needs. For founders, agent loops multiply token counts, so cheap reasoning tokens on hardware the customer already owns can change unit economics and ease data-residency objections. The benchmarks are vendor-run, and the open-weight field is crowded this week, so test on your own tool-call traces before committing.