Deploy local agents everywhere with LFM2.5-2.6B

LFM2.5-2.6B is a model designed to enable powerful on-device agents. It supports tool calling and multi-step workflows while being compact enough to run on everyday hardware like laptops and phones. This allows developers to deploy agents without relying on cloud inference, keeping data private and reducing costs. The model is pre-trained on 34 trillion tokens, with a mid-training phase that extends the context window to 128,000 tokens.

After training, it is transformed into an agent through four stages: optimization, inference, and environment execution. The model has been evaluated against larger models and has shown competitive performance in instruction following, tool use, and agentic tasks. It outperforms other models in instruction following and tool use, with only one exception in tool use benchmarks.

In agentic tasks, it matches or exceeds performance compared to other models. It also excels in knowledge and performs similarly in math. However, larger models still lead in coding tasks. The model is compatible with various inference ecosystems, including llama.cpp, MLX, vLLM, SGLang, and ONNX. It also supports CPU inference, making it possible to run capable agents on mobile devices.

GPU inference allows for almost 15,000 output tokens per second at high concurrency, which could process around 1.3 billion tokens per day on a single H100. LFM2.5-2.6B is the fastest model in its class and is ideal for high-volume workloads. The model is available on Hugging Face and can be installed with the latest version of transformers. Both LFM2.5-2.6B and LFM2.5-2.6B-Base are now available for use.