NVIDIA's Nemotron for Science hits 61% on quantum dot synthesis, beating GPT-6 Astra at 10-20x lower cost
ryanshrout · x · 2026-11-17
At SC26, NVIDIA pitched AI agents taking on more scientific work with announcements covering the full stack, from open models to a CPU-only simulation rack and a new HPC GPU.
- Nemotron for Science pairs the open Nemotron 3.5 Super VL model with a 50,000-question chemistry dataset and post-training recipes. Lila Sciences' tuned version hit 61% accuracy on quantum dot synthesis vs 54% for GPT-6 Astra at 10-20x lower inference cost; KTH used a Nemotron agent to fix 81% of broken GROMACS setups overnight on a single DGX Spark. Open models labs can tune and self-host address cost and data-control concerns that keep institutions off closed frontier APIs. Argonne and RIKEN are among users.
- Vera is flagged by the author as the most consequential near-term release (text truncated).
Used by Argonne, RIKEN and more.
More from Infra
- Accelerator forecasts bundle memory; sell-side still sees over $1T market by 2030 — BenBajarin · 2026-11-17
- Ben Bajarin: $9.6T AI capex vs $4.1T tracked revenue — but raising paid fleet share to 50% closes the 2030 bill — BenBajarin · 2026-11-17
- Ex-NVIDIA Research Director recalls 12.5 years, calls recent week 'orchestrated and targeted' — PavloMolchanov · 2026-11-17
- Modal Sandboxes burst instead of reserving: half of sandbox life uses under 6% peak CPU — AAAzzam · 2026-11-17
- Mintlify cuts search latency 35% after migrating to turbopuffer — handotdev · 2026-11-17
- Infrawrench CI claims 2–5× faster than GitHub Actions with WASM sandboxes — threepointone · 2026-11-17