Periodic Labs introduces Periodic Neon, a 1T-parameter lab-trained model that beats GPT-6 Astra at X-ray diffraction analysis
Event confirmedAwaiting review
The takeaway
On Sept 15, 2026 Periodic Labs, the materials-science startup of Liam Fedus and Ekin Dogus Cubuk, published its first research.
Status
- Claim
Event confirmedAwaiting review
- Our reporting
- High confidence
- Verification
- Unverified claim
- Importance
- 3 of 5
- Last verified
- 12 October 2026
Your AI and this story
- GPT-6.2 Astra107 days after its cutoff
- Claude Opus 5.577 days after its cutoff
- Gemini 4 Hydrogen168 days after its cutoff
- Grok 4.7107 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 77 days before it.
Key facts
- 1T parameters; post-trained from open-weight Kimi K2.6 with scientific midtraining + RL on proprietary XRD lab data
- Final training run on 1,300 H200 GPUs
- FrontierXRD (internal eval): Neon 55.3% vs Kimi K2.6 2.7%; company says it surpasses GPT-6 Astra and Claude Fable 5.1 ‘at a lower cost per analysis’ (their scores not given in the post)
- Deployed in Periodic’s physical labs to analyse experiments in the search for superconductors and magnets; no discoveries claimed yet
- Three posts the same day: ‘Building Labs that Learn’, ‘Nature Is Our Learning Environment’, ‘AI Infrastructure at Periodic’
- No announcement of releasing Neon’s weights
- Liam Fedus on The Frontier podcast with Mercor’s Brendan Foody (reported by Synced via 36Kr, Oct 12): ‘Machine learning systems are not magic’, AlphaFold worked because decades of protein-structure data existed; published materials results are a biased sample because ‘everything you see on the Internet is a success’; Periodic runs its own Menlo Park lab because ‘without the ability to put ideas into practice, you can only stay at the thinking stage’ (quotes as translated by 36Kr)
What happened
Periodic Labs came out of stealth in 2025 with a $300M seed round. Its first public results show the “lab as RL environment” idea: experimental data from its own autonomous labs is used to specialise a large open-weight model for one scientific task, analysing XRD patterns.
Why it matters
It shows a startup taking a Chinese open-weight frontier model and, with modest compute and proprietary experimental data, beating the best closed models on a narrow scientific task. That is an argument that experimental data, not general capability, is the bottleneck in AI-for-science.
What is disputed or not yet verified
| Verification | Company-reported on an internal benchmark; not independently replicated |
|---|
Sources
5 sources from 3 sites. Numbers match the chips in the text.
5 sources: 3 primary, 2 press
Primary
- Periodic Labs: Building Labs that Learnperiodic.com, official
- Periodic Labs: Nature Is Our Learning Environmentperiodic.com, official
- Periodic Labs: AI Infrastructure at Periodicperiodic.com, official
Press
Changes
- Added Fedus’s podcast remarks (36Kr)
- Filed