A trillion-parameter Periodic Neon moves into the physical lab: post-trained on in-house diffraction data, it beats Claude Fable 5.1 and GPT-6 Astra on analysis
Synopsis
Periodic Neon is a trillion-parameter model post-trained with midtraining and reinforcement learning on data from the company's own labs and deployed to its high-throughput physical labs in Menlo Park to analyze X-ray diffraction results. It outperforms Claude Fable 5.1 and GPT-6 Astra on diffraction analysis, and inference is optimized so the model can be deployed at scale. Its training data comes from physical experiments rather than elastically scalable digital environments.
Interpretation
A trillion-parameter model has been post-trained and deployed into physical laboratories to analyze experimental results, notably X-ray diffraction (XRD) patterns, with the aim of putting autonomous discovery into scientists' hands. XRD analysis historically required hours of scientific judgment drawing on experimental conditions, simulations and prior results; that step is now handled by a model trained on lab data. On diffraction analysis, Periodic Neon outperforms both Claude Fable 5.1 and GPT-6 Astra, using relatively little compute.
The model's frontier-level reasoning comes from midtraining followed by reinforcement learning, with training data drawn from experiments the labs themselves produce. Reinforcement learning in digital environments relies on elastically scalable numbers of agents and fast, verifiable tasks; physical experiments cannot scale agent count freely because power, equipment and engineering are needed, each run can take days, and results are often ambiguous. Training and deployment target the high-throughput labs in Menlo Park, which run 24/7 alongside infrastructure for large-scale physics simulations aimed at new materials such as superconductors and magnets.
Perspective
The demonstrated scope is diffraction analysis in the Menlo Park high-throughput labs, together with midtraining and reinforcement learning on that lab's data; deployment targets optimized inference at scale. The team says it is extending the approach to directing scientific campaigns, developing synthesis procedures and deciding which experiments to run.
The number of agents for physical experiments is not elastically scalable, a single experiment can take days, and results are often ambiguous; these are conditions the approach must face. The team expects that as labs and compute scale, AI trained on lab data will tackle increasingly complex scientific problems, which is an expectation rather than a reported result.