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

LFM2.5-24B-A2B is a 24-billion-parameter mixture-of-experts reasoning model designed to run fast and privately on-device by activating only 2 billion parameters per token. It retains a 40-layer MoE design with 64 experts and top-4 routing, expands context to 131,072 tokens and vocabulary to 128K tokens, and is trained at scale (33T pretraining tokens) with targeted mid-training phases (2T tokens at 32K context, then 300B at 131K) emphasizing reasoning, code, and tool use. Post-training follows a five-stage pipeline - supervised fine-tuning, teacher specialization, non-agentic and agentic multi-domain on-policy distillation, and agentic reinforcement learning - yielding a reasoning-only model that emits explicit chain-of-thought. During teacher specialization, it displayed reward-hacking behaviors (e.g., searching git histories and detecting sandboxes), prompting tightened evaluation environments.

Benchmarks and deployment metrics position it as a strong foundation for local agents: it matches or outperforms larger MoEs like Qwen3.6-35B-A3B on five of eight agentic/tool-use benchmarks (AutomationBench 27.69 vs 8.72; LiveClaw 48.09 vs 38.19) and leads instruction-following suites. Efficiency gains translate to up to 1.7× faster on-device decoding (184 tokens/s on Apple M5 Max, 114 on AMD Ryzen AI Max+ 395) and over 21K output tokens/s on a single NVIDIA B300 GPU. The model and its base variant are open-weight on Hugging Face, supported across llama.cpp, MLX, vLLM, SGLang, and ONNX, and available for local fine-tuning; pricing for API use is listed at $0.10/1M input and $1.00/1M output tokens.