Liquid AI Releases LFM2.5-24B-A2B for Fast On-Device Agents
A 24B-parameter reasoning model with only 2B active per token, beating larger MoEs on five of eight agentic benchmarks while decoding at 184 tokens/s on an Apple M5 Max.
Liquid AI has released LFM2.5-24B-A2B, the largest and final member of the LFM2.5 family, which now spans from 230M to 24B parameters. It activates only 2B parameters per token, and unlike its predecessor it is reasoning-only, emitting an explicit chain of thought before answering.
The upgrades over LFM2-24B-A2B are substantial: context extended from 32K to 131,072 tokens, vocabulary doubled from 65K to 128K, and pretraining scaled from 17T to 33T tokens. Post-training runs in five stages, ending with agentic reinforcement learning inside real harnesses including OpenClaw, Hermes Agent, Pi, and OpenCode.
On benchmarks it leads or matches Qwen3.6-35B-A3B — a model with roughly 50% more total and active parameters — on five of eight tool-use and agentic evaluations, scoring 48.09 on LiveClaw (vs. 38.19) and 27.69 on AutomationBench (vs. 8.72). It also leads all three instruction-following benchmarks, though larger models retain the edge on knowledge-heavy tests like GPQA Diamond and on SWE-Bench Verified.
Speed is the real argument: 184 tokens/s on an Apple M5 Max, up to 1.7× faster than comparable MoEs, and over 21K output tokens/s on a single NVIDIA B300. Open weights are on Hugging Face with day-one llama.cpp, MLX, vLLM, SGLang, and ONNX support.