Deploy local agents everywhere with LFM2.5-2.6B
Hugging Face released LFM2.5-2.6B, a compact language model designed to run agentic tasks entirely on local devices, enabling private, cloud-free deployments for developers.
Hugging Face introduced LFM2.5-2.6B, a 2.6-billion-parameter model optimized for on-device agentic workflows, including tool use and multi-step tasks. The model is pre-trained on approximately 34 trillion tokens with a 128K context window and post-trained through a four-stage reinforcement learning pipeline. This architecture separates model optimization, inference, and environment execution, allowing agents to operate without cloud dependency while maintaining data privacy. The design supports deployment on laptops, phones, and other edge devices, reducing infrastructure costs for developers.
Benchmark tests compared LFM2.5-2.6B against models up to four times its size across STEM, instruction following, tool use, and agentic tasks. Despite its smaller size, it outperformed larger models in instruction following and tool use, with the exception of BFCLv4, where a 9.7-billion-parameter Qwen model led. On agentic tasks, it matched or exceeded performance of Gemma and Qwen variants, though larger models retained an advantage in coding tasks. The model demonstrates strong capabilities in knowledge and math benchmarks, positioning it as a practical choice for edge deployments.
LFM2.5-2.6B achieves high inference speeds on both CPU and GPU, making it suitable for real-time agentic applications. On an M5 Max CPU, it decodes at 220 tokens per second, while a Ryzen AI Max+ 395 handles 113 tokens per second. On GPU, it reaches nearly 15,000 output tokens per second at high concurrency, enabling high-volume workloads on a single H100. These performance metrics allow developers to deploy capable agents even on mobile devices, supporting scalable, low-latency use cases without cloud costs.
The model is available today on Hugging Face, with both LFM2.5-2.6B and its base variant accessible for immediate use. Day-one support is provided across major inference frameworks, including llama.cpp, MLX, vLLM, SGLang, and ONNX. A browser-based demo demonstrates the model powering a research agent that answers questions and generates summaries, showcasing its practical applications for developers building on-device agentic systems.