Vera Rubin sweeps MLPerf on debut
NVIDIA's Vera Rubin platform has arrived on the industry's standard yardstick. In results released 16 November, it swept every MLPerf Training v6.1 benchmark and led its first MLPerf Endpoints submission for agentic workloads. The new Vera Rubin NVL72 system delivered up to twice the performance per GPU of a highly optimized GB300 NVL72, and more than 10× the throughput per watt on agentic inference — including a provisional result on the open Kimi K3 model at the same response speed. On DeepSeek-V3 training, it scaled from 256 to 1,024 GPUs at about 97% efficiency, cutting the run from 16 minutes to about four. NVIDIA's report.
Nemotron for Science, with open chemistry recipes
NVIDIA also introduced Nemotron for Science, pairing its open Nemotron 3.5 Super VL model with a new chemistry dataset, LitmusBench v0.2 evaluations, and post-training recipes for domain-specific science agents. The model and LitmusBench dataset are on Hugging Face; NVIDIA says Super VL is optimized for a 128GB DGX Spark, and Argonne is among the early users. NVIDIA.
OpenAI opens ChatGPT to 200,000 researchers
OpenAI is expanding its ChatGPT for Academic Researchers programme to give up to 200,000 researchers worldwide free access. The scheme began as a July pilot; applications for the next round are now opening. OpenAI.
Computer-use agents for healthcare paperwork
Separately, OpenAI published work on computer-use agents for healthcare teams, aimed at administrative workflows such as assembling what an insurer needs after a clinician orders an MRI. It names Abridge, Ambience, Cigna Healthcare and CVS Health among the organisations it is working with. OpenAI.



