Periodic Labs' Neon beats GPT-6 Astra using just 1,300 H200s and a lab data loop
LiamFedus · x · 2026-09-16
Liam Fedus (ex-OpenAI VP of research, now Periodic Labs) announced Neon, an open-source model mid-trained and RL'd with only 1,300 H200s plus months of proprietary experimental data from their high-throughput materials lab in Menlo Park, surpassing GPT-6 Astra on their materials-analysis benchmark. Focus areas: superconductors, magnets, and semiconductor materials.
His takeaways:
- Fine-tuning open models to reliably beat flagship closed models on specific tasks isn't trivial; naive fine-tuning can look good on benchmarks but remain brittle
- Periodic's success combined deep ML, systems work, domain expertise, and a lab-model data flywheel
- For most enterprises, he recommends GEPA-style optimization or waiting for stronger base models before investing in custom training
More from Models
- LightOnOCR-3 Reads 16th-Century Manuscripts and Ancient Greek Without Training — IgorCarron · 2026-11-02
- 'You Can Feel the Gradient Slopes': Dev Contrasts Benchmaxxed RL Models vs General Ones — theteknosaur · 2026-11-02
- LightOnOCR-3 Launches: Tops FRBench, Adds Layout Structure and Chart Extraction — IgorCarron · 2026-11-02
- NVIDIA's Kumo Is a Family of Four Tabular Foundation Models, Not Just One — pandeyparul · 2026-11-02
- SemiAnalysis: safety classifiers are underrated — Meta's Muse tops cyber benchmark at 35.2% — ThomasScialom · 2026-11-02
- DeepSeek V4.1 Flash 1017 wins over a longtime closed-source-only user: fast, cheap, reliable — _bachrc · 2026-11-02