Liquid AI announced the availability of the LFM2.5-2.6B and LFM2.5-2.6B-Base models on Hugging Face, enabling AI agents to run locally on devices rather than relying entirely on cloud inference. According to the company, the model supports tool calling and multi-step workflows, with a size and speed suitable for use on laptops and phones.
The release targets developers who need to deploy agents at scale while keeping data on the device and reducing dependence on a cloud inference bill. Liquid AI says the model competes with models approximately four times larger in tool use, instruction following, and multi-step agent tasks.
Training Tailored to Agent Tasks
LFM2.5-2.6B was pretrained on approximately 34 trillion tokens, and the mid-training phase included expanding the context window to 128,000 tokens. The model then went through four stages to transform it from a base model into an agent capable of handling tools and sequential tasks.
- Supervised fine-tuning: Two rounds of training focused heavily on tool-use data, web search, and agent workflows.
- Teacher specialization: Training a specialized teacher model for each domain, such as mathematics, programming, and tool use.
- Multi-domain distillation: Transferring the capabilities of specialized models into a single model.
- Agentic reinforcement: Running multi-round training within actual agent frameworks so the model could learn to handle different tools, system prompts, and multi-step task environments.
The agentic reinforcement system separates model improvement, inference, and environment execution. The system uses the training engine to improve the model and the rollout engine to generate actions using the latest policy, while the reinforcement framework coordinates training operations and the collection of trajectories and rewards. Actions are executed inside a Sandbox service, where the Blackbox Harness hosts the agent and coordinates its interaction with the task environment, without requiring modifications to the supported frameworks.
Test Results and Performance
Liquid AI compared the model with models approximately four times its size, including gemma-4-E2B-it, gemma-4-E4B-it, Qwen3.5-4B, and Qwen3.5-9B. LFM2.5-2.6B was the smallest model in the group, but it topped all instruction-following tests included in the comparison. It also topped the tool-use tests except for BFCLv4, in which Qwen3.5-9B outperformed it.
The model scored 80.07 on Multi-IF, 85.49 on IFStruct, 77.83 on ToolSandbox, and 62.85 on the English Claw-Eval average. On agent tasks, it outperformed the two Gemma models and maintained performance comparable to the Qwen models, while the larger models remained clearly ahead in programming.
Runtime Speed and Availability
From day one, the model is compatible with several inference tools, including llama.cpp, MLX, vLLM, SGLang, and ONNX. The company says generation speed reaches 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395 processor, with the model running in less than 2.5 gigabytes of memory. It also indicates that phone-capable agents can run at a speed of 30 tokens per second.
On graphics processing units, the model reaches approximately 15,000 output tokens per second at high concurrency, equivalent to roughly 1.3 billion tokens per day on a single H100. Developers can try the WebGPU demo in the browser or use the model with local agent frameworks such as OpenClaw, Hermes Agent, and Pi. A demo of an on-device research agent is also available.