NVIDIA opened four Cosmos 3 Nano checkpoints on 5 November 2026 — post-trained world models of 16 billion trainable parameters each, for robot and self-driving simulation, under the OpenMDW 1.1 licence (commercial and non-commercial use permitted).
The four models cover different simulation jobs. Transfer-Auto generates multi-view driving footage — up to 11 camera views including fisheyes, with joint LiDAR output — with varied weather and lighting. Sim-Auto streams camera-and-LiDAR observations in real time. Sim-Transfer turns physics-simulator controls into photorealistic video. Sim-Bimanual predicts video from a two-arm robot's actions.
NVIDIA also published the full four-stage post-training recipe and coding-agent skills so teams can adapt the models to their own sensors and robots. Generation needs about 58 GiB of GPU memory in BF16, 44 GiB in FP8, and 31 GiB with FP8 offloading (NVIDIA's own figures). No independent team has yet measured the checkpoints on real robotics or driving workloads.
Key facts
| Item | Details |
|---|---|
| Release | 5 November 2026 |
| Models | Four Cosmos 3 Nano checkpoints, 16 billion trainable parameters each |
| Purpose | Robot and self-driving simulation |
| Licence | OpenMDW 1.1 (commercial and non-commercial use permitted) |
| Transfer-Auto | Multi-view driving footage — up to 11 camera views including fisheyes, joint LiDAR output, varied weather and lighting |
| Sim-Auto | Real-time camera-and-LiDAR observation streams |
| Sim-Transfer | Physics-simulator controls into photorealistic video |
| Sim-Bimanual | Video prediction from a two-arm robot's actions |
| Also published | Full four-stage post-training recipe and coding-agent skills |
| GPU memory (NVIDIA's figures) | About 58 GiB (BF16), 44 GiB (FP8), 31 GiB (FP8 offloading) |
What this means
Simulation is where robots and self-driving systems learn most of what they know — real-world testing is slow, expensive and sometimes dangerous. Open-weight world models that can generate realistic driving footage and robot observations lower the cost of building that training data, and the OpenMDW 1.1 licence permits commercial use.
The practical bar is hardware: generation needs tens of gigabytes of GPU memory, so these are models for teams with serious GPUs, not for a laptop. And until an independent team measures the checkpoints on real workloads, NVIDIA's own figures are the only performance numbers available — treat them as the vendor's figures.
A note on verification
The release details are from Hugging Face (nvidia/Cosmos3-Nano-FP8; HF blog), pondero.ai, the64th.com, gpujournal.ai and dev.to, following a Quora lead by AI Tech Tools. The lead is credited in the source line; all content here is written in our own words. The 5 November date is just outside this pass's two-week window; the release is included because it is a significant open-weights release.
FAQs
What is Cosmos 3 Nano?
Four open-weight world models from NVIDIA, opened 5 November 2026, of 16 billion trainable parameters each, for robot and self-driving simulation.
What do the four models do?
Transfer-Auto generates multi-view driving footage; Sim-Auto streams camera-and-LiDAR observations in real time; Sim-Transfer turns physics-simulator controls into photorealistic video; Sim-Bimanual predicts video from a two-arm robot's actions.
What licence do the models use?
OpenMDW 1.1, which permits commercial and non-commercial use.
How much GPU memory does generation need?
About 58 GiB in BF16, 44 GiB in FP8, and 31 GiB with FP8 offloading — NVIDIA's own figures. No independent team has yet measured the checkpoints on real workloads.
Source: NVIDIA release, 5 November 2026 (Hugging Face nvidia/Cosmos3-Nano-FP8 and HF blog; pondero.ai; the64th.com; gpujournal.ai; dev.to); Quora lead: AI Tech Tools (https://www.quora.com/What-is-your-opinion-on-Nvidias-Cosmos-foundation-models-that-help-train-robots-and-self-driving-cars/answer/AI-Tech-Tools-1). Last checked: 21 November 2026.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!