High-throughput materials labs in Menlo Park provided the private data used to train a specialized AI model called Neon for superconductor analysis. Periodic Labs mid-trained a Kimi 2.5 base model to 1 trillion parameters using 1,300 H200 GPUs, surpassing GPT-6 Astra and Fable on a materials analysis benchmark. The combination of this training and reinforcement learning increased X-ray diffraction analysis success from 2.7% to 55.3% across 134 samples.

The model targets hard problems in superconductors, magnets, and semiconductor materials using a process termed reinforcement learning from world feedback. This loop allows Neon to help run physical experiments and subsequently learn from the results. Infrastructure optimizations for the workload produced 4.1 times the training throughput of a Megatron baseline and 2.5 times faster decoding with over 95% cluster utilization.

Periodic Labs contributes its technical improvements back to the open-source community through projects including Megatron-LM, SGLang, and Miles. The project also included the development of pbox, a proprietary sandbox solution designed specifically for scientific workloads.