Deploy local agents everywhere with LFM2.5-2.6B

Details

Liquid AI has released LFM2.5-2.6B, designed to bring capable multi-step AI agents to everyday hardware without cloud dependency. Pre-trained on approximately 34 trillion tokens with a 128K context window, the model achieves 220 tokens/second on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU while fitting under 2.5 GB of memory. The post-training pipeline is what distinguishes this model for agentic tasks. After two rounds of supervised fine-tuning weighted toward tool use and harness trajectories, Liquid trained specialist teacher models per domain (math, code, tool use), then distilled them into the 2.6B student via multi-domain on-policy distillation. A final Agentic Reinforcement Learning stage ran multi-turn RL inside real harnesses like OpenClaw and Hermes Agent. Benchmark results show LFM2.5-2.6B outperforming Gemma 4 (5.1B and 8B) and Qwen3.5-4B on AIME25 (51.87 vs 26.33/34.27/49.33) and IFBench (59.17 vs 34.x). It is the smallest model in its tested group while holding competitive positions on agentic tasks. For developers, this means a privacy-preserving, zero-cloud-cost inference agent on a laptop or phone — relevant wherever data sovereignty, latency, or usage cost are constraints.