Extropic Boosts Qwen3.6 Thermo ML Performance 2.8x in 100 Steps

Extropic improved the Qwen3.6-35B-A3B model's performance on thermodynamic ML tasks by 2.8x using reinforcement learning on Prime Intellect infrastructure, raising its reward from 0.127 to 0.361 after approximately 100 GRPO steps. The post-trained model, which activates only 3B parameters per token, outperformed other open models and narrowed the reward gap with Claude Opus 4.8, with similar gains reported for the Qwen3.5 model family. Extropic developed a custom RL environment with verifiers to accelerate algorithmic discovery and used Prime Intellect's managed services to avoid managing multi-node GPU infrastructure.