The authors release LFM2.5-2.6B, a 2.6B parameter agentic model designed to run entirely on-device for planning, tool calling, and multi-step tasks. It is small enough to operate on phones and fast enough to remain responsive on CPUs, offering free inference, low latency, and privacy without cloud API dependencies.

  • The model was pre-trained on ~34T tokens and features a 128K vocabulary extended for non-Latin scripts.
  • Post-training utilizes a four-stage pipeline: Supervised Fine-Tuning, Teacher Specialization, Multi-Domain On-Policy Distillation, and Agentic Reinforcement Learning.
  • It leads on instruction-following benchmarks and outperforms Gemma models on agentic tasks, while trailing Qwen3.5-9B only on specific math and coding metrics.
  • Inference is supported via llama.cpp, MLX, vLLM, SGLang, and ONNX, achieving up to 15K output tokens per second on an NVIDIA H100.

This enables developers to build massively parallelized agents that run around the clock on local hardware at no marginal cost.