On November 5, 2026, NVIDIA released four open-weight checkpoints in the Cosmos 3 Nano series via Hugging Face, formally extending interactive world simulation capabilities to the autonomous driving and robotics tracks. The four checkpoints are Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer, and Cosmos3-Nano-Sim-Bimanual. Each model contains 16 billion trainable parameters and is released under the OpenMDW-1.1 license governed by the Linux Foundation, allowing commercial and non-commercial use and requiring the "Built on NVIDIA Cosmos" attribution.
This release is the largest expansion of the Cosmos 3 family since its main launch at GTC Taipei/Computex on May 31, 2026. At that time, NVIDIA CEO Jensen Huang positioned Cosmos 3 in his keynote as an "open frontier foundation model for physical AI" and simultaneously released weights in two sizes: Super with 64 billion parameters and Nano with 16 billion parameters. The 4-billion-parameter Edge version followed in July of the same year.
Architecture Logic: From a Multi-Model Pipeline to Single-Model Full Coverage
To understand the value of these four new checkpoints, it is necessary to first clarify what Cosmos 3's underlying architecture does. Technical analysis shows that earlier physical AI systems typically relied on chained calls: one model handled scene understanding, another handled world generation, and a third predicted robot actions, with integration losses at every interface. Cosmos 3 replaces this pipeline with a Mixture-of-Transformers dual-tower architecture.
The division of labor between the two towers is as follows: the Reasoner Tower, as an autoregressive vision-language model, handles understanding and reasoning; the Generator Tower, as a diffusion Transformer, generates physics-aware video and future frames. The two share 3D rotary position embeddings, and the Generator must run on top of the Reasoner's context output. Inference preceding generation is a hard design constraint. The entire system was trained on approximately 20 trillion tokens of image, video, audio, and action data.
The four checkpoints released on November 5 are all versions post-trained on Nano for specific tasks, following a four-stage training scheme of "bidirectional adaptation → teacher forcing → consistency distillation → self-forcing." This scheme focuses on addressing historical error accumulation in autoregressive generation. This means developers receive not a general-purpose base model, but specialized models already tuned for target scenarios and directly pluggable into simulation pipelines.
Four Checkpoints, Two Industry Tracks
Functionally, these four models clearly cover two application layers:
On the autonomous driving side, Transfer-Auto generates multi-view driving data covering different weather, lighting, and times of day; Sim-Auto outputs synchronized camera images and LiDAR point clouds that can be directly connected to driving simulators; Sim-Transfer converts simulator control signals and text prompts into photorealistic video, addressing the domain shift between simulated scenarios and real-world distributions.
On the robotics side, Sim-Bimanual supports interactive simulation of YAM and AgiBot dual-arm robot embodiments, and can be controlled by policy models or operated in real time by a person via a controller. This interface design means researchers can collect dual-arm manipulation data without actually deploying hardware, or use human teleoperation to fill edge-case training sets.
Such open, task-specific world-model checkpoints reduce the cost of generating rare driving scenarios and robot training data, eliminating the need to collect each extreme condition one by one in reality. This is a direct extension of the logic of using simulation in place of real data, and it is especially important for autonomous driving teams: sudden cut-in scenarios on snowy nights can be generated in large quantities in simulation without waiting to encounter them by chance in real road testing.
The Hardware Reality Behind the Open License
Cosmos 3's licensing strategy is relatively friendly by current large-model ecosystem standards: OpenMDW-1.1 allows commercial use and derivative models and requires only attribution. Compared with the Llama series' restrictions on commercial scale or the fully closed strategies of some models, these licensing terms do leave room for startups and research institutions.
The hardware threshold is another matter. Technical guidance shows that Nano, with 16 billion parameters, requires at least 96GB of VRAM, corresponding to workstation-class RTX PRO 6000 setups or multi-GPU configurations; Super, with 64 billion parameters, must run on H100, H200, or B200 data-center GPUs. Edge, with 4 billion parameters, is intended for embedded deployment with lower hardware requirements and targets Jetson devices, but its covered capabilities are correspondingly narrower.
This means the accessibility of "open weights" is clearly stratified in practice: the infrastructure of academic teams and large autonomous driving companies meets the requirements, but most robotics startups still face a barrier to running Nano inference directly on local hardware. Cloud inference can bypass VRAM constraints, and NVIDIA provides the Cosmos NIM microservice interface for this purpose, supporting GPU-accelerated hosted calls.
Ecosystem Aggregation and the Signal of 10M Downloads
In terms of development pace, official NVIDIA forum data shows that the Cosmos series surpassed 10 million downloads by July 27, 2026. Accumulating this volume in less than eight months indicates that demand from the developer community for physical AI infrastructure genuinely exists and is expanding rapidly.
NVIDIA is also advancing the "Cosmos Coalition," seeking to integrate world-model builders, AI developers, and companies in the physical AI field into an ecosystem alliance. Robotics companies such as Figure AI and Boston Dynamics are exploring the use of Cosmos 3 for world-model training for humanoid robots and autonomous driving. This ecosystem strategy is highly similar to the CUDA path of the past: build developer dependence with an open toolchain, then use hardware to pull the entire value chain back.
NVIDIA's strategic intent in this release is clear: expand the ecosystem with open weights, lower access costs with task-specific checkpoints, and use hardware tiers to guide enterprise users toward data-center configurations. This is a structurally complete ecosystem layout, but the real value of physical AI must ultimately be delivered on robots and vehicles under real-world operating conditions.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接