Periodic Neon ('Nature Is Our Learning Environment')

Summary

Periodic Neon is a 1-trillion-parameter scientific model created by midtraining on Kimi K2.6 plus reinforcement learning on data from Periodic's own physical labs, deployed to analyze experiments in the search for better superconductors and magnets. On the internal FrontierXRD evaluation (134 hard X-ray diffraction analysis samples), it reaches 55.3% success — a 20x improvement over Kimi K2.6's 2.7% — and is claimed Pareto-optimal on cost vs performance against GPT-6 Astra and Claude Fable 5.1.

Keywords

scientific AI · materials discovery · XRD · mid-training · reinforcement learning · autonomous labs

Topics

scientific AI, materials discovery, XRD, mid-training, reinforcement learning

Research notes

• Method: Scientific midtraining + RL on lab data using 1,300 H200 GPUs; evaluated with the Periodic Harness and an LLM-judge ensemble (Opus 5, GPT-5.6-Sol) calibrated against expert ratings.
• Key findings: 55.3% on FrontierXRD vs 2.7% for base Kimi K2.6 (20x improvement); Claimed to beat GPT-6 Astra and Claude Fable 5.1 at lower cost per analysis on this task; Generalized to held-out chemical systems not seen in midtraining/RL
• Limitations: All results are Periodic's own reported numbers on an internal eval; third-party analysis notes the 'beats frontier at science' headline needs a narrow task-specific reading.
• Details from the Periodic post excerpt plus Import AI #474 and TLDR coverage via web search; post page itself was rate-limited.