Periodic Labs' Neon took X-ray diffraction analysis from 2.7% to 55.3% by training inside its own lab.

Periodic mid-trained and post-trained a 1T-parameter model on roughly 1,300 H200s using months of fresh data generated by its own high-throughput materials labs, closing a model-experiment loop rather than learning from a static corpus. Liam Fedus reported the jump across 134 samples and the team claims it beats GPT-6 Astra on their analysis benchmark. The structural claim for anyone building domain agents: the proprietary bottleneck was physical experiment throughput, not compute or base model quality.