Liquid AI has released LFM2.5-2.6B, a compact language model designed for deploying local agents. The model emphasizes efficient operation on edge devices, enabling developers to build and run AI agents in local environments.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Liquid AI releases LFM2.5-2.6B, a small language model designed for on-device agents, rivaling models four times its size in tool calling and multi-step tasks while maintaining low memory usage and high inference speed.
LFM2.5-2.6B is a 2.6B parameter language model released by Liquid AI on August 4, 2026, designed specifically for on-device agents. It supports tool calling and multi-step workflows, running on everyday hardware like laptops and phones, enabling developers to deploy agents on-device, protecting data privacy and avoiding cloud inference costs.
According to the official blog, the model competes with models four times its size in tool use, instruction following, and multi-step agent tasks, and is called the 'best-in-class agent.' Its pre-training data amounts to approximately 34T tokens, with the context window extended to 128K through intermediate training.
The post-training process is divided into four stages: two rounds of supervised fine-tuning (SFT), with emphasis on agent data; teacher specialization, training expert teachers for domains like math, code, and tool use; multi-domain policy distillation (MOPD), distilling expert teachers into a single student model; and agentic reinforcement learning (Agentic RL), conducting multi-round reinforcement learning in real agent frameworks.
The Agentic RL pipeline separates model optimization, inference, and environment execution, coordinating through a training engine, replay engine, and sandbox service. Black-box frameworks like OpenClaw or Hermes Agent can be integrated without modification, while capturing token-level trajectories for training sample reconstruction and validation.
Official benchmarks show LFM2.5-2.6B excels in instruction following and tool use. It scores 59.17 on IFBench, 80.07 on Multi-IF, and 85.49 on IFStruct, all surpassing comparison models. In tool use, it scores 77.83 on ToolSandbox and 56.88 on BFCLv4, second only to Qwen3.5-9B.
On agent tasks, it averages 62.85 on Claw-Eval, comparable to the Qwen series but better than Gemma models. On the math benchmark AIME25, it scores 51.87, close to Qwen3.5-4B's 49.33 but below Qwen3.5-9B's 56.07. On the code benchmark LiveCodeBenchv6, it scores 59.41, trailing all comparison models, with the official acknowledgment that code capability is a weakness.
Thanks to the LFM2 architecture, LFM2.5-2.6B achieves decoding speeds of 220 tok/s on CPU (Apple M5 Max) and 113 tok/s (AMD Ryzen AI Max+ 395), with memory usage under 2.5 GB. On GPU, output speed approaches 15K tokens per second under high concurrency, with a single H100 processing about 1.3B tokens per day.
On the first day of release, the model supports inference frameworks like llama.cpp, MLX, vLLM, SGLang, and ONNX, and provides a browser WebGPU demo. Developers can load the model via the transformers library, compatible with transformers>=5.0.0.
LFM2.5-2.6B and LFM2.5-2.6B-Base are available for download on Hugging Face. The official provides guides for running local agents in frameworks like OpenClaw, Hermes Agent, and Pi.
The model aims to realize the vision of 'AI running anywhere,' providing a solution for high-throughput on-device agent workloads.
This report is based on Liquid AI's official blog; all benchmark data are official self-reported results and have not been independently verified. Performance may vary with hardware and configuration.
LFM2.5-2.6B demonstrates the potential of small models in on-device agent scenarios, but code capability is weak, and benchmark data require independent verification.
Primary report
Primary source