NVIDIA Open-Sourced Four 16B World Models. The Part Nobody Benchmarked Is Policy Learning.

On November 5, NVIDIA put four Cosmos 3 checkpoints on Hugging Face. Each is a post-trained version of the 16 billion parameter Cosmos 3 Nano, published under OpenMDW1.1, a licence that permits commercial and non-commercial use. Alongside the weights, the company released the post-training recipes and the inference code, so a team can adapt the models to its own sensor rig rather than only run them as shipped.
Then, the next day, NVIDIA shipped VFlash, a decoding framework that attacks a different part of the same problem. The two releases are more interesting together than apart, because one is about generating the world and the other is about how fast you can say where things are in it.
What the four checkpoints do
Cosmos 3 is a world foundation model family that launched in June with a Mixture-of-Transformers backbone: an autoregressive subsequence handling reasoning and a diffusion subsequence handling generation, with separate encoders per modality and a shared representation space in the middle. Two sizes exist, Super and Nano. The November release is four specialised post-trains of the smaller one.
Cosmos3-Nano-Transfer-Auto generates offline driving data. It produces synchronised video across up to 11 camera views, fisheyes included, plus LiDAR point clouds. Road layout and traffic behaviour are set per camera through a World Scenario Map, and appearance is steered by text.

Cosmos3-Nano-Sim-Auto is the causally distilled streaming version of Transfer-Auto. It emits small chunks of synchronised camera video and LiDAR in a few denoising steps, which is what you need for long interactive rollouts instead of one fixed clip.
Cosmos3-Nano-Sim-Transfer takes a physics simulator's controls plus a text prompt and returns photorealistic RGB video. The physics state stays in the simulator, and the model is conditioned on four spatial controls: depth, segmentation, edges and blur. That division is the useful bit. You keep the dynamics you trust and borrow the pixels you don't have.
Cosmos3-Nano-Sim-Bimanual predicts the next video frame from bimanual robot or camera actions. It supports the YAM and AgiBot embodiments, and a single control sequence can switch between viewpoint changes and manipulation. The mixed-action example in the documentation uses 59 control values.
The training detail worth reading
NVIDIA published the four-stage recipe: bidirectional adaptation, teacher forcing, consistency distillation, and self-forcing.
The last one is the one to pay attention to. Self-forcing trains the model on its own generated history rather than only on ground truth frames, which is the standard trick for slowing error accumulation in autoregressive video. If you train only on clean context, the model never learns to recover from its own drift, and the drift compounds once it is running free.
Anyone who has watched AI video fall apart over ten seconds has seen the symptom. Publishing the fix in recipe form, with weights and code, is more useful than publishing another demo reel.
The sentence that should be quoted more
NVIDIA says it remains to be measured whether these models reliably support policy learning and evaluation.
That is an unusually honest line for a launch, and it should travel with every story about open world models. What the company is saying is that a world model which produces believable video has not thereby been shown to be a useful training environment. A policy trained inside a simulator that renders beautifully but is physically wrong will learn the wrong thing and fail in the real world, and the failure will look like a hallucination rather than a physics bug.
Closing that gap requires evaluation that most teams are not set up to run: take a policy trained in the generated world, deploy it on real hardware, and measure the delta. That is expensive, slow, and not the kind of result that fits in a launch post.
So the honest framing of this release is that NVIDIA has handed the community four strong generators plus the recipes, and explicitly declined to claim they work as training environments. The claim is the community's to earn.
VFlash, and why coordinates are not words
Autoregressive decoding was designed for language, where the next token is a reasonable unit. Geometry is not like that. A robot needs a grasp point, an orientation, and a path, and generating those one coordinate at a time is both slow and fragile. Each prediction conditions on the previous one, so an early error propagates. Ten waypoints means ten chances to drift.
VFlash replaces token-by-token generation with what NVIDIA calls flash primitive decoding. The model declares how long the output will be through a tool call, declares the primitive type, and then a lightweight draft model predicts the whole geometric object in parallel while the target vision language model stays frozen. Three engineering problems have to be solved for that to work: boundary alignment, so a decode block maps to one complete geometry rather than being sliced by a token window; position alignment, so coordinates and waypoints always land in fixed-width fields; and geometry structure learning, which is what the draft model is for.
The draft model is a five-layer non-causal Transformer trained while the target VLM is completely frozen. Text subtasks still go through speculative decoding. The whole thing is a plug-in, not a retrain.
The numbers are specific. On an H100 at batch size one, 3D point motion prediction reaches up to 25 times the single-generation speed. Robot action decoding is up to 12 times faster. End-to-end robot inference is up to 5.1 times faster. 3D trajectory error drops by as much as 61 percent, with average 3D displacement error down 42.6 percent across six metrics on ShareRobot-Bench and VABench-V. On DAVIS, dense 3D motion prediction cuts displacement error by 27 percent. Bounding box generation on COCO and Dense200 runs 2.6 to 2.8 times faster at roughly baseline accuracy.
And in MolmoSpaces, a world action model conditioned on VFlash visual planning reaches 70.2 percent average task success against 62.6 percent for the baseline.
Jim Fan's summary was that the era of the System 1 vision model has arrived, and that the next inference revolution is physical reasoning at reflex speed. Marketing, but the direction is right. Reflex-speed control needs decoders that emit geometry in one shot.
What a team should take from the pair
If you are building anything that touches the physical world, the practical reading is that the two halves of the stack just got cheaper.
The generation half: four 16B checkpoints you can download, with per-camera scenario control, LiDAR synchronisation, and a simulator-conditioned mode that keeps physics where you can trust it. Requirements are unglamorous: Linux, an NVIDIA GPU, Docker, and the container toolkit.
The control half: a decoding framework that turns a frozen VLM into something fast enough to plan at control rates, without retraining it.
The half nobody shipped is the evaluation. Whether a policy trained on Cosmos-generated data transfers to real hardware is still an open question, and NVIDIA says so. That is the work that will decide whether open world models are a training-data shortcut or a very pretty renderer.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.