Skip to content
NewsModel release

NVIDIA opens four Cosmos 3 Nano checkpoints for robot and self-driving simulation

· by

The short version

NVIDIA released four open-weight Cosmos 3 Nano checkpoints on November 5, 2026, with full post-training recipes that let robotics and autonomous-vehicle teams adapt the models to their own sensors rather than just run a demo.

NVIDIA opens four Cosmos 3 Nano checkpoints for robot and self-driving simulation

NVIDIA opened four checkpoints on November 5, 2026 that predict what a camera or lidar sensor would see next in a simulated scene, in real time, and published the full post-training recipe needed to adapt them to a team's own sensor rig, not just the weights, per NVIDIA's Hugging Face announcement.

What

The four checkpoints, Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer, and Cosmos3-Nano-Sim-Bimanual, extend NVIDIA's Cosmos 3 model family. Per the Hugging Face post, each handles a different slice of simulation: generating multi-view driving footage with varied weather and lighting, streaming synchronized camera-and-lidar observations, converting a physics simulator's controls into photorealistic video, and predicting video from robot or camera actions. All four are post-trained versions of Cosmos3-Nano with 16 billion trainable parameters, released under NVIDIA's OpenMDW 1.1 license, which the company says permits both commercial and non-commercial use. NVIDIA frames the split along two use cases in the Hugging Face post: the two autonomous-vehicle checkpoints take a text prompt plus World Scenario Map controls (lanes, objects, road markings, traffic signs, road boundaries) so a team can vary road layouts and signal states and watch how a driving policy reacts, while the robotics checkpoints turn simulator controls or robot and camera actions directly into predicted video, which the company says helps answer whether a policy still works once the scene or viewpoint changes. Three days earlier, on November 2, NVIDIA published a companion technical post describing the broader Cosmos Framework that trains and serves these checkpoints, including a GPU memory table showing Cosmos3-Nano generation needs 58 GiB of VRAM in BF16, dropping to 44 GiB in FP8 and 31 GiB with FP8 offloading. The same table lists a larger Cosmos3-Super variant at up to 150 GiB in BF16, so the Nano checkpoints are the entry point, not the ceiling, of the family. The same post says speculative decoding in NVIDIA's Cosmos NIM microservices measured roughly a 2 to 3 times inference speedup on H200 GPUs, and that streaming video understanding built on the same checkpoints can cut alert latency by a factor of 10 for live camera feeds.

Teams get the training recipe, not just a demo

What separates this release from a typical model drop is the four-stage recipe NVIDIA published alongside the weights. Per the Hugging Face post, training runs through bidirectional adaptation on complete recorded clips, then teacher forcing, then consistency distillation down to four denoising steps, and finally a self-forcing stage where the model trains on its own generated history so small errors do not compound across frames. NVIDIA published the code for each stage as agent-runnable skills under the Cosmos Framework GitHub repo, one skill folder per checkpoint, so a robotics or autonomous-vehicle team can point a coding agent at its own dataset and retrain rather than settle for NVIDIA's example footage. That is a meaningfully lower bar than a closed API: a team still needs real GPU budget (the memory table above sets the floor), but it does not need to rebuild the training pipeline from a paper.

The framework also exposes reinforcement-learning paths NVIDIA says are new in this release, per the November 2 technical post: offline preference optimization (DPO), online RL with reward models and vLLM rollouts (GRPO), actor-critic RL for embodied policies in simulation (PPO), on-policy distillation from a frozen teacher (OPD), and an online RL variant built for diffusion world models (DiffusionNFT). NVIDIA says the checkpoints run on Hugging Face Diffusers, PyTorch, SGLang, vLLM, vLLM-Omni, TRT-LLM, and TRT-Edge-LLM, which keeps a team's choice of serving stack separate from its choice of training recipe.

Context and reactions

The release sits inside a larger expansion NVIDIA describes in the November 2 technical post: Cosmos Framework now extends beyond Cosmos 3 to cover NVIDIA Alpamayo, GR00T, Qwen, and pi0.5, positioning the training and inference tooling as shared infrastructure across robotics and autonomous-vehicle projects rather than a one-off research release. NVIDIA's own engineers were already using an early Cosmos3-Nano checkpoint internally before the wider release. A separate November 5 NVIDIA blog post on agent-assisted simulation workflows describes an NVIDIA autonomous-vehicle researcher running what the post calls "a separate Cosmos3-Nano experiment" that varied weather and lighting in recorded driving footage to compare how an Alpamayo driving model responded under different conditions, per NVIDIA's "Into the Omniverse" post. That internal use case predates and is distinct from the Cosmos 3 checkpoint release itself, but it points to how NVIDIA expects the models to get used day to day: as a tool inside an existing simulation pipeline, directed by an AI agent, rather than a standalone product.

What to watch next

NVIDIA's GPU memory table also lists Cosmos3-Edge and Cosmos3-Super variants that have not shipped open checkpoints yet; whether those follow the same Nano-first pattern is worth tracking. NVIDIA scheduled a Cosmos Labs livestream for November 5 to walk through the framework, and the replay is posted on YouTube for teams evaluating whether to adopt it. The real test will be whether outside robotics and AV teams publish their own post-trained checkpoints on top of these recipes in the coming weeks, which is the clearest signal an open world-model release actually gets used rather than just downloaded once.

Sources