Liquid AI released LFM2.5-2.6B on Monday, a 2.6-billion-parameter model designed to run capable AI agents entirely on consumer hardware. The model is available on Hugging Face with open weights and targets on-device deployment across laptops and phones.
LFM2.5-2.6B achieves 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen CPU, fitting within 2.5 GB of memory. Liquid AI positions it as competitive with models roughly four times its size on tool use, instruction following, and multi-step agentic tasks.
Benchmarks
On instruction-following benchmarks, the model tops every category against comparable models including Gemma 4 E2B (5.1B), Qwen 3.5 4B (4.7B), and Gemma 4 E4B (8B). It scores 59.17 on IFBench and 80.07 on Multi-IF, well ahead of the next best 5.1B model at 34.08 and 69.44 respectively. On BFCLv4, a tool-calling benchmark, it reaches 56.88, behind only Qwen 3.5 9B at 60.13.
On math, LFM2.5-2.6B scores 51.87 on AIME25, competitive with Qwen 3.5 4B at 49.33. On LiveCodeBench v6 it scores 59.41, ahead of both Gemma models.
Training pipeline
Liquid AI used a four-stage post-training pipeline to build the model's agent capabilities. After pre-training on roughly 34 trillion tokens and extending the context window to 128K tokens, the team applied supervised fine-tuning with an emphasis on tool-use and agent trajectory data, trained specialist teachers per domain, distilled them into a single student via multi-domain on-policy distillation, and finally ran multi-turn reinforcement learning inside real agent harnesses.
The agentic RL pipeline separates model optimization, inference, and environment execution into distinct components. A training engine optimizes the model while a rollout engine generates actions using the latest policy. Actions execute inside a sandboxed black-box harness where the model interacts with task environments. A harness proxy captures token-level trajectories for RL training without modifying the harness itself.
The model was trained inside popular agent harnesses to improve compatibility across different tools, system prompts, and multi-turn task environments.
