Nvidia posted four new open-weight Cosmos 3 Nano checkpoints on November 5, 2026, aimed at interactive world simulation for driving scenes and two-arm robot video prediction. The Hugging Face blog describes each checkpoint as a post-trained Cosmos 3 Nano example with 16 billion trainable parameters under OpenMDW 1.1 terms that permit commercial and non-commercial use subject to the license. The release expands Nvidia’s simulation stack beyond closed demos and sits squarely in The Boston Report’s AI Hardware coverage alongside prior chip and export reporting such as Nvidia RTX Pro 5500 China order expectations.
What each checkpoint is for
According to Nvidia’s Hugging Face article, Cosmos3-Nano-Transfer-Auto generates synchronized multi-view driving video (up to 11 cameras, including fisheyes) with optional joint RGB–LiDAR for offline synthetic data. Cosmos3-Nano-Sim-Auto is a distilled streaming variant for chunk-by-chunk camera and LiDAR rollouts. Cosmos3-Nano-Sim-Transfer maps simulator controls and a text prompt into photorealistic RGB video. Cosmos3-Nano-Sim-Bimanual predicts the next visual observation from an image plus bimanual robot or camera actions, with support described for YAM and AgiBot embodiments. Those descriptions are Nvidia’s; independent policy-learning benchmarks for the new weights were not published in the launch post.
By the numbers
All four checkpoints are described as 16B-parameter post-trained Nano models. The chart below treats that as four equal open releases rather than a performance ranking. OpenMDW 1.1 is Nvidia’s stated license family for these weights. Runtime notes on Hugging Face mention Linux, an Nvidia GPU, Docker and the Nvidia Container Toolkit, with a published vLLM-Omni setup tested on one GB200 — Nvidia says that is not a minimum-memory specification. Related developer framing also appears in Nvidia’s technical blog on building world models with Cosmos Framework.
Hardware and software stack implications
Open simulation weights matter for teams that want to generate rare driving weather, lighting and multi-sensor layouts without collecting every case on the road, and for labs testing two-arm manipulation policies before physical trials. They do not replace calibrated physics engines for contact-rich tasks, and Nvidia’s own materials frame the checkpoints as starting points for domain adaptation. For semiconductor and AI infrastructure readers, the release is a software/hardware platform story more than a new GPU SKU — but it still routes through Nvidia’s GPU software ecosystem.
How this fits TBR’s AI hardware desk
TBR’s hardware hub filters Artificial Intelligence posts tagged to semiconductors, Nvidia or export controls. This checkpoint release is tagged accordingly so it surfaces on /ai-news/hardware/ without editing hub queries. Broader AI model posts remain on the models landing.
What developers still have to measure
Nvidia’s launch post is careful about evaluation: the company presents the checkpoints as recipes and examples for adaptation rather than as finished, universally scored driving or manipulation agents. That means fleet operators and robotics labs still need their own closed-loop metrics — collision rates, route completion, grasp success — before treating generated video as a substitute for physical data. Hugging Face comments and secondary write-ups likewise stress that usefulness for training and evaluating policies remains to be measured case by case.
For hardware buyers, the practical question is GPU memory and throughput when running 16B-class world models interactively. Nvidia’s published GB200 test note is a configuration example, not a sizing guide for every shop. Teams without Blackwell-class hardware should assume they will need to distill, quantize or offload before interactive rates are usable. Those engineering costs belong in any budget that treats “open weights” as free.
Export-control readers should note that open simulation weights are not the same product class as restricted data-center GPUs, but they still sit inside Nvidia’s broader stack. TBR tags this piece for the hardware hub because the story is about Nvidia’s computing platform for embodied AI, not because it announces a new SKU price.
Sources
Read next in Artificial Intelligence
Most read this week
- Abdul El-Sayed and the Muslim Brotherhood Claims: What’s Confirmed and What Isn’t
- Stock Futures Erase a 2.4% Overnight Drop as Markets Digest Democrats’ Sweep of Congress
- Spain Transport Minister Óscar Puente Quits 26 Days Before Election Over Crash That Killed 69
- GM Korea’s Sales Jump 26.6% on Chevrolet Exports as Hyundai Slips 6.3% and Korean Factory Growth Cools
- Boston Plans to Close 3 Schools in 2028 as Enrollment Drops as Much as 58%
Latest
- Google Backstops TeraWulf Debt for Anthropic’s $19B Kentucky Data Center Lease
- AU Chair Meets Abiy as Sudan File and Horn Crises Crowd Agenda
- Von der Leyen Says EU Coordinated G7 Oil Stock Releases With IEA
- Macquarie Profit Jumps 50% to A$2.48B on Commodities Surge
- Mélenchon Platform and Le Pen Align Against NATO Command