Guava open-sourced to turn frontier AI into compact robot agents

Guava’s developers say they open-sourced a framework for distilling frontier models into compact robot agents using only simulation interaction data. They say they fine-tuned Qwen3.5 into Guava-4B with about 2,000 simulation trajectories and report 90.0% real-world success, versus 93.3% for GPT-5.4, with 7.1× lower model-call latency through local inference.