Periodic Labs' Neon: trillion-param science model beats GPT-6 Astra with 1,300 H200s
LiamFedus · x · 2026-09-16
Periodic Labs (co-founded by ex-OpenAI's Liam Fedus) unveiled Neon, an open-source trillion-parameter model trained on data from its own high-throughput materials labs in Menlo Park, closing the loop between experiments and models: labs generate fresh data, models learn from it and suggest what to try next.
Key facts:
- Using only a peak of 1,300 H200s for mid-training and RL, Neon surpasses GPT-6 Astra and Claude Fable 5.1 on their X-ray diffraction benchmark, targeting superconductors, magnets, and semiconductor materials.
- Built on Megatron, SGLang, Miles, and Ray, heavily modified for scientific RL where rollouts reason and run tools for hours while training steps take minutes.
- 4.1x training throughput over standard stacks, SOTA memory efficiency and long-context training (long scientific traces stress memory and parallelism), plus 95%+ cluster utilization by sharing capacity with simulation workloads.
The pitch: real experimental data plus extreme engineering efficiency lets a small cluster beat frontier models in a specialized scientific domain.
More from Infra
- Chris Lattner to keynote PyTorch Conference on Mojo and MAX for heterogeneous compute — PyTorch · 2026-09-16
- US data centers may use more natural gas than Germany and Japan combined by 2035 — TechCrunch AI · 2026-09-16
- Four US locales pause data-center builds in one week; Google, Amazon projects exempt — BenBajarin · 2026-09-16
- Researchers Gained Admin Access to Baseten's Production GitHub in 25 Minutes via PAT Takeover — bearsyankees · 2026-09-16
- Event-driven multi-agent architecture beat model routing on cost, 32-agent finance system shows — tradfyAi · 2026-09-16
- China narrows memory chip gap with Korea: 3 years in HBM, 1 year in NAND — robleclerc · 2026-09-16