source: Hugging Face Blog: Deploy local agents everywhere with LFM2.5-2.6B
level: technical
Liquid AI released LFM2.5-2.6B, a 2.6-billion-parameter language model designed to run agentic workflows entirely on local hardware. It supports tool calling and multi-step tasks on laptops and phones, keeping data private and avoiding cloud costs. The model was pre-trained on roughly 34 trillion tokens and fine-tuned with supervised learning, multi-domain distillation, and agentic reinforcement learning inside real agent harnesses like OpenClaw.
On instruction-following benchmarks, LFM2.5-2.6B scored 59.17 on IFBench and 80.07 on Multi-IF, surpassing Gemma and Qwen models up to 9.7B parameters. For tool use, it reached 56.88 on BFCLv4 and 77.83 on ToolSandbox, trailing only the larger Qwen3.5-9B on BFCLv4. It achieves 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen CPU, using under 2.5 GB of memory.
The model is available on Hugging Face with support for llama.cpp, MLX, vLLM, and ONNX. It targets developers who need private, cost-effective agents for high-volume tasks. While it leads in instruction following and tool use, larger models still hold an edge in pure coding benchmarks. The release includes a browser demo and guides for integrating with agent frameworks like Hermes Agent and Pi.
why it matters: It enables private, low-cost agent deployment on consumer devices, reducing reliance on cloud APIs for tool-using AI applications.
source: Hugging Face Blog: Deploy local agents everywhere with LFM2.5-2.6B