Periodic Labs' Neon Trained on 1,300 H200s Outperforms GPT-6 Astra in Materials Science

Periodic Labs, co-founded by former OpenAI post-training lead Liam Fedus, published "Building Labs that Learn," detailing its materials science model Neon. The team built a 24/7 high-throughput materials lab in Menlo Park, forming a closed loop where the lab produces new data, the model learns from it, and the model guides the next experiments across superconductors, magnets, and semiconductors. Using only 1,300 H200 GPUs plus months of proprietary experimental data, with mid-training and RL on an open-weight base, Neon Pareto-dominates frontier models like GPT-6 Astra on XRD benchmarks.

Confirmed

• The blog lays out three pillars—high-throughput lab, research, and infrastructure—arguing that frontier labs waste vast amounts of data, and that a closed loop can continuously strengthen scientific AI
• Neon is a 1-trillion-parameter model trained to analyze experimental data, already deployed in the company's own lab for superconductor, magnet, and semiconductor experiments
• Training started from open weights with mid-training and RL stages, peak compute of only 1,300 H200s; on the 134-sample internal Frontier set, XRD analysis success rose from 2.7% to 55.3%, roughly 20x
• XRD analysis normally takes scientists hours of cross-referencing software, literature, and databases; Neon automates much of it
• Team member Ollmer's infrastructure blog describes how robust ML infrastructure with multimodal, lab-grounded RL extends AI's scientific capability, summarized as "nature is the ultimate test environment"
• Liam Fedus confirmed close collaboration with the radixark team (SGLang project), which supported Neon's infrastructure and research, and emphasized open-source community contributions
• Diederik Bahdanau relayed that Periodic Labs is completing the positive loop of "AI runs the lab, AI learns from the lab"
• xeophon cited the XRD post-training results as among the most impressive OpenAI-ecosystem-adjacent outputs, as evidence of closed-lab investment in scientific data

Unconfirmed

• Some reposters (e.g., in teortaxesTex's thread) claimed the base model is Kimi 2.5; this comes from discussion rather than official disclosure, and the specific base model remains unconfirmed, as does whether the model will actually be open-weight

Why it matters

• Demonstrates that proprietary experimental data plus an integrated closed loop can beat frontier general models on specific scientific tasks with far less compute, offering a replicable playbook for vertical scientific AI
• Automated labs and models reinforcing each other could materially accelerate discovery of superconductors, magnets, and semiconductors
• If the loop of "AI runs the lab, AI learns from the lab" works at scale, it could become new infrastructure for AI for Science

2026-09-16 ~ 2026-09-17 · 69 related posts

• [source] Periodic Labs trains Neon on just 1,300 H200s to beat GPT-6 Astra on materials analysis — LiamFedus · 2026-09-16
• [source] Periodic Labs unveils "labs that learn": 1T-param Neon now deployed in its own labs — LiamFedus · 2026-09-16
• [source] Periodic Neon lifts XRD analysis success from 2.7% to 55.3% — a 20x jump — LiamFedus · 2026-09-16
• Periodic Labs' Neon, trained on 1,300 H200s, beats GPT-6 Astra on its science benchmark — _AndrewZhao · 2026-09-16
• Periodic Labs trains Neon with just 1,300 H200s, beating GPT-6 Astra on materials analysis — LiamFedus · 2026-09-16
• Periodic Labs: 1,300 H200s beat frontier models on X-ray evals, 4.1x Megatron throughput — zephyr_z9 · 2026-09-16
• Liam Fedus' Lab Built Neon: 1,300 H200s and a Self-Driving Materials Loop Beat GPT-6 Astra at Superconductor Analysis — vwxyzjn · 2026-09-16
• Open-source model Neon mid-trained on 1,300 H200s claims to beat GPT-6 Astra — bookwormengr · 2026-09-16
• Periodic Labs unveils 1T-parameter Neon model that beats GPT-6 Astra on diffraction analysis — agarwl_ · 2026-09-16
• Periodic Labs trains a 1T-parameter XRD analysis expert on its lab data — lishali88 · 2026-09-16
• OpenAI's Liam Fedus unveils Neon: scientific models trained with modest compute — UltraRareAF · 2026-09-16
• Periodic Labs midtrains and RLs a trillion-param LLM to analyze superconductor lab data, beating Astra and Fable — vwxyzjn · 2026-09-16
• Ex-OpenAI post-training lead Liam Fedus confirms move into physical-world AI — LiamFedus · 2026-09-16
• Periodic Labs pushes Kimi 2.5 base model past Astra with specialized scientific training — teortaxesTex · 2026-09-16
• Richard Sutton endorses Periodic Labs' Neon, tuned on 1,300 H200s — LiamFedus · 2026-09-16
• SGLang team radixark powered the infra behind Periodic Labs' Neon — LiamFedus · 2026-09-16
• Neon cost math questioned: $2.5/H200-hour internal pricing vs Kimi K2.5 API — eliebakouch · 2026-09-16
• Periodic Labs shows specialized scientific models can rival frontier AI on modest compute — LiamFedus · 2026-09-16
• Periodic Labs reportedly trained an open-weight model on 1,300 H200s that beats GPT-6 Astra internally — hsu_byron · 2026-09-16

31 near-duplicate retellings: vwxyzjn · LiamFedus · LiamFedus · vwxyzjn · LiamFedus · LiamFedus · LiamFedus · vwxyzjn · LiamFedus · vwxyzjn · LiamFedus · LiamFedus · sudoraohacker · BanghuaZ · ying11231 · LiamFedus · willdepue · andrew_n_carr · LiamFedus · LiamFedus · giffmana · eliebakouch · LiamFedus · LiamFedus · vwxyzjn · LiamFedus · LiamFedus · YuchenHe07 · LiamFedus · vwxyzjn · LiamFedus