NVIDIA Open-Sources Cosmos 3 World Models for Real-Time Interactive Physical Simulation

NVIDIA Open-Sources Cosmos 3 World Models for Real-Time Interactive Physical Simulation

On November 5, 2026, NVIDIA published the Cosmos 3 family of world foundation models on Hugging Face, releasing seven checkpoints across Predict 3, Transfer 3 and Reason 3, with the largest model at 12B parameters capable of action-conditioned rollouts at 720p and 30fps for up to 30 seconds per run. It is the first time a general-purpose world model arrives with commercially usable weights and on-premise deployment, which could loosen the data bottleneck holding back physical AI. The core

NVIDIA posted the Cosmos 3 release to its Hugging Face blog at 19:35 on November 5, 2026, with weights simultaneously listed on build.nvidia.com and the NGC catalog. Seven checkpoints shipped: Cosmos-Predict 3 in 4B and 12B sizes for text-, image- and camera-pose-conditioned video rollouts; Cosmos-Transfer 3 at 7B for mapping multimodal control signals onto photorealistic frames; and Cosmos-Reason 3, an 8B vision-language model handling physical commonsense and action planning.

The decisive change from the original Cosmos line shown at CES in January 2025 is interactivity. The models accept joint angles, end-effector displacement and six-degrees-of-freedom camera poses as input sequences, then run closed-loop rollouts at 720p and 30fps, extending single-run duration from roughly five seconds to 30 seconds, with a 20-second physical-consistency window stated on the model card. Training

Why now. The world-model race has moved from demos to engineering over the past two years: DeepMind's Genie 3 showed a 720p, 24fps interactive environment in August 2025 but opened it only to selected researchers, while World Labs and Decart sell playable worlds as products. The open-source side has lacked general weights with action conditioning, leaving teams to either build simulators themselves or rely on physics-engine approaches like Isaac Sim, which are precise but visually thin and expensive to author assets for.

NVIDIA's position makes openness the rational move. As a compute supplier its returns come from downstream training and inference volume, not model licensing; the first

Liu Gong's take: the watershed for Cosmos 3 is not image quality but the action-conditioning interface, which turns a world model from a video generator into a simulator that a policy network can call. Within 12 months, robotics teams will have their data pipelines rewritten before the content-creation industry notices; closer in, there is a strong chance that within six months vertical simulation vendors appear, selling fine-tuned, commercially licensed interactive worlds for specific industrial scenarios.

From a technical standpoint, the combination of action conditioning and a 30-second consistency window is the substantive number here. The first Cosmos generation produced non-interactive clips, so a robot could not learn that doing A leads to B. The third generation takes joint angles and camera poses as input and extends rollouts to 30 seconds, which happens to cover the full cycle of grasping, insertion and tightening tasks. In industrial inspection scenarios, Liu Gong's observation is that the bottleneck in synthetic data was never sharpness but causal consistency of defects: move the light source and does the scratch stay at the same coordinate, does the shadow follow it. If Transfer 3 projects geometric constraints from control signals faithfully onto pixels, the cost of collecting samples at every line changeover drops materially, which matters far more than attractive demo renders.

On the business side, this is the classic pattern of the shovel seller paving the road first. At the same 720p tier, Genie 3's 24fps is open to a handful of researchers, while Cosmos 3 offers 30fps under a commercial license; the gap is availability, not parameter count. One is a paper, the other is a component. NVIDIA gives up licensing revenue on the model to drive adoption of Isaac Lab, Omniverse and GB200, because the more free the model, the more compute buyers need to run it. CUDA already proved this playbook once.

For the industry, the real variable is evaluation. When synthetic data can be generated without limit, the dividing line shifts from whether you have enough data to whether your simulation is faithful enough, and whoever owns the metric for sim-to-real consistency owns pricing power. Open weights accelerate that, and accelerate commoditization just as fast.

  • Training data: the Cosmos 3 model card states 30 million hours of video, versus 20 million hours disclosed for the first Cosmos generation in January 2025, a roughly 50% increase.
  • Interaction length: single-run rollout for the 12B model rises from about five seconds to 30 seconds, roughly six times longer, with a stated 20-second physical-consistency window.
  • Frame rate and resolution: Cosmos 3 runs at 720p/30fps, against the 720p/24fps that DeepMind disclosed for Genie 3 in August 2025, about 20% lower frame rate, with no open weights.
  • Parameter spread: Predict 3 in 4B and 12B, Transfer 3 at 7B, Reason 3 at 8B, seven checkpoints total, all under the commercially usable OpenMDW 1.1

For the industry, the near-term shift within three months is straightforward: robotics teams begin wiring Cosmos 3 into existing synthetic-data pipelines, replacing part of the asset-authoring work currently done in graphics engines, and the composition of VLA training sets visibly changes. Within a year, vertical interactive-world vendors take shape, and sim-to-real consistency metrics become a core clause in procurement negotiations.

For ordinary users, little is visible in three months; the effect hides inside the failure rates of delivery robots and warehouse sorting arms. Within a year, real-time responsive generative scene tools reach content creators, though consumer-facing products

Questions worth exploring:

  • How action-conditioned interfaces change the cost structure of collecting robot training data
  • Open weights with commercial licensing as NVIDIA's compute lock-in strategy for physical AI
  • Evaluation standards for interactive world models: how to measure sim-to-real consistency

Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.

This item comes from Hugging Face Blog (published 2026-11-05 19:35:16). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.


Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.

Republishing: feel free to quote or repost with a link back to the original article on Wenfeng Radar. · Business & partnership: see our media kit
Daily briefing + Sunday review · Unsubscribe anytime by replying · Your email stays private

Leave a Reply

Your email address will not be published. Required fields are marked *

中文EN
Contact · Privacy · Disclaimer · AI Disclosure · Weekly Review · Media & Business
Email: wenfengleida@wenfengleida.cn