Liquid AI releases on-device agent model

Liquid AI launched LFM2.5-2.6B, a 2.6 billion-parameter model designed to act as a local agent that can plan actions, call tools and execute multi-step tasks entirely on the device, without cloud API calls. The company calls the version “open-weight”; the training data and the training code have not been released.

The model was trained on roughly 34 trillion tokens, with a vocabulary of 128 thousand tokens and a context window of 128 thousand tokens. Fine-tuning proceeded in stages: separate expert models were refined for mathematics, code, tool use and long-context handling; their capabilities were then distilled into a single model via on-policy distillation; the final stage took place in real agent environments using Hermes Agent and OpenClaw.

According to Liquid AI, the model generates up to 220 tokens per second on an Apple M5 Max, about 113 tokens per second on a Ryzen AI Max+ 395, and roughly 30 tokens per second on a smartphone. On a single H100 under high parallel load the output rate reaches about 15 thousand tokens per second. CPU memory usage stays below 2.5 GB, allowing comfortable execution on modest hardware.

In the company’s benchmarks, LFM2.5-2.6B outperforms larger models from the Gemma and Qwen families on instruction and tool-use tasks, while the larger models retain an edge on complex coding. The model is supported by llama.cpp, MLX, vLLM, SGLang and ONNX, with both base and instruction variants available for download.