Periodic Labs' Neon Beats GPT-6 Astra and Claude Fable 5.1 on Materials Benchmark

Periodic Labs has released Periodic Neon, a one-trillion-parameter model fine-tuned from the open-weight Kimi K2.6 architecture, which surpasses GPT-6 Astra and Claude Fable 5.1 on the FrontierXRD materials analysis benchmark. On that benchmark, Neon achieved a 55.3% success rate on complex samples compared to 2.7% for the unmodified K2.6 baseline. The model was trained using 1,300 H200 GPUs and months of proprietary experimental data generated by Periodic Labs' physical high-throughput materials labs in Menlo Park, run by a team including Liam Fedus. The core system works as a closed loop: the physical labs generate fresh experimental data, models learn from it, and models then guide which experiments to run next.