NVIDIA shipped Cosmos 3 on June 17, 2026 through its official NVIDIA developer hub and a public Hugging Face collection. The release includes three open-weight physical AI models: Cosmos 3 Nano 4B, Cosmos 3 Pro 14B, and Cosmos 3 Ultra 34B. Nano and Pro use Apache 2.0. Ultra uses the NVIDIA Open Model License. The family targets robotics, autonomous driving simulation, and embodied agent research. This is a physical AI stack, not a text model. It predicts video, depth, and low-level robot actions from real sensor inputs. The models are available as safetensors with no gated access for the 4B and 14B versions.
Why it matters is simple. Closed physical AI models from OpenAI, Google DeepMind, and Tesla dominate commercial robotics. Research teams get API access but cannot inspect weights or fine-tune on private robot data. Cosmos 3 changes that for many teams. The 14B model handles 32,768 token context. The 34B model reaches 131,072 tokens, enough for long video scenes. NVIDIA’s own Nemotron 3 Ultra 550B release showed the company’s open-weight push. Cosmos 3 extends that push to physical AI. You can now train a world model without paying per token, which matters as AI free tier limits tighten across major providers.
Release specifics matter for planning. The 4B Nano model runs on a single RTX 4090. The 14B Pro model needs one 80GB GPU such as an A100. The 34B Ultra model is aimed at H100 or DGX clusters. Context windows are 8,192, 32,768, and 131,072 tokens. Nano and Pro are Apache 2.0, while Ultra uses the NVIDIA Open Model License with restrictions on at-scale autonomous vehicle fleets. This is more permissive than many closed APIs but not completely unrestricted. It follows the approach seen with NVIDIA’s RTX Spark Superchip edge AI line. Teams that need local inference can avoid monthly tool pricing.
Early developer reaction focuses on practical access. The weights are on Hugging Face and NVIDIA NGC. A GitHub organization includes inference scripts, model configs, and example datasets. The open model compares to closed physical AI simulators from Waymo and Tesla, but independent audits are still early. Cosmos 3 should not be treated as a production safety system. It is a research and development tool that lowers the cost of physical AI experiments. For free model access and local deployment options, see best open-source LLM models 2026. The immediate effect is that smaller robotics teams now have an open starting point that did not exist at this quality level before.
How Do the Top Options Compare?
| Model or Access | Parameters | Context Window | License | Best For |
|---|---|---|---|---|
| Cosmos 3 Nano 4B | 4 billion | 8,192 tokens | Apache 2.0 | Edge robots and single GPU |
| Cosmos 3 Pro 14B | 14 billion | 32,768 tokens | Apache 2.0 | Simulation and fine-tuning |
| Cosmos 3 Ultra 34B | 34 billion | 131,072 tokens | NVIDIA Open Model License | AV simulation and large fleets |
| Hugging Face Community Inference | Depends on model | Depends on model | Apache 2.0 / NOML | Quick testing without GPUs |
| NVIDIA NGC SDK | Depends on model | Depends on model | Apache 2.0 / NOML | Enterprise deployment with Isaac |
Nano and Pro use Apache 2.0. Ultra uses the NVIDIA Open Model License with restrictions on at-scale autonomous vehicle fleets. FP16 sizes are about 8 GB for Nano, 28 GB for Pro, and 68 GB for Ultra. Community inference and NGC are access paths, not separate model architectures.
1. Cosmos 3 Nano 4B , Best for edge robots and single-GPU local testing
Cosmos 3 Nano is the smallest model in the June 17, 2026 release. It has 4 billion parameters and an 8,192 token context window. The model processes RGB, depth, and IMU data to generate short video predictions and low-level robot actions. NVIDIA reports that Nano runs at roughly 12.4 frames per second at 512x512 on an RTX 4090. With 4-bit quantization, it fits in about 4.5 GB of VRAM, which makes it usable on many developer laptops and edge devices. You can download weights from Hugging Face or the NVIDIA developer hub. This model is ideal for rapid prototyping, teleoperation replay, and pick-and-place tasks.
Nano uses an Apache 2.0 license, so commercial use, modification, and redistribution are allowed. That is a major advantage over closed robot APIs that charge per inference or restrict weight access. The main tradeoff is shorter context. You cannot feed it a long driving video or multi-minute indoor navigation sequence without truncation. Still, for single-arm manipulation and small mobile robots, it is often enough. If you are new to local physical AI, start with the Hugging Face free inference path before buying a dedicated GPU.
Key strengths:
- ✅ Runs on a single consumer GPU
- ✅ Apache 2.0 commercial use
- ✅ Small 4.5 GB quantized footprint
- ✅ Good for short teleoperation replay tasks
- ❌ 8k context limits long video scenes
- ❌ Less physics detail than Pro and Ultra
Who it’s for: Robotics hobbyists and edge teams that need a cheap, local physical AI model.
2. Cosmos 3 Pro 14B , Best for simulation, fine-tuning, and mid-size robotics fleets
Cosmos 3 Pro is the middle option for research teams that need stronger multi-step physics prediction. It has 14 billion parameters and a 32,768 token context window. NVIDIA reports a PhysBench-CoT score of 43.2 for Pro, up from 31.8 on the previous Cosmos generation. The model predicts longer action sequences and maintains object permanence better than Nano. In simulation tests, Pro reduced trajectory drift by 29 percent compared with Nano on a 60-second manipulation benchmark. This makes it the most likely starting point for labs that fine-tune on private robot data. You can download Pro from the same Hugging Face collection. FP16 inference requires about 28 GB of VRAM, so an A100 80GB or H100 80GB is recommended.
Pro uses an Apache 2.0 license, which is rare for a model at this capability level in physical AI. You can fine-tune, merge, or ship it in commercial products without per-device royalties. The model is also supported by the broader open-source ecosystem. The state of open-source on Hugging Face spring 2026 report shows physical AI repos growing faster than language model repos. If your team already runs local LLMs, Pro fits into a similar self-hosted pipeline. For general local model context, see best open-source LLM models 2026 coding local agentic AI.
Key strengths:
- ✅ Stronger multi-step physics prediction
- ✅ 32k context handles longer video
- ✅ Apache 2.0 permits fine-tuning and commercial use
- ✅ Balanced hardware demand for labs
- ❌ Requires an 80GB GPU for FP16
- ❌ 4-bit quantization reduces physics fidelity
Who it’s for: Research labs and robotics startups that need open fine-tuning for complex manipulation.
3. Cosmos 3 Ultra 34B , Best for high-fidelity video world models and autonomous vehicle simulation
Cosmos 3 Ultra is the largest open-weight physical AI model in the family. It has 34 billion parameters and a 131,072 token context window. That context length supports long driving scenes, multi-room robot navigation, and continuous video generation for world models. NVIDIA reports a 37 percent reduction in collision rate compared with Cosmos 2.5 in its internal CARLA traffic simulation. Ultra also scores highest on multi-view depth consistency and long-horizon object tracking benchmarks. The model is designed for H100 or DGX deployments. FP16 weights take roughly 68 GB of VRAM, so you need at least one 80GB GPU or a sharded multi-GPU setup.
Ultra is not completely unrestricted. It uses the NVIDIA Open Model License, which permits research and most commercial use but restricts at-scale autonomous vehicle fleets above a defined threshold. That is a meaningful limitation for robotaxi companies. For most robotics labs and simulation providers, the license is workable. It is similar in spirit to NVIDIA’s Nemotron 3 Ultra 550B open weight strategy: open enough to build on, but with guardrails at the top of the market. If you need closed-scale physical AI alternatives, the cost conversation is covered in AI free tier limits June 2026.
Key strengths:
- ✅ 131k context for long video
- ✅ Best physics fidelity in the family
- ✅ Optimized for NVIDIA H100 clusters
- ❌ Heavy GPU requirements
- ❌ NVIDIA Open Model License restricts large AV fleets
Who it’s for: Well-funded labs and simulation providers that need the best open physical AI model.
4. Hugging Face Community Inference , Best for quick testing without local GPUs
For developers who do not have a high-end GPU, Hugging Face hosts community inference and demo spaces for Cosmos 3 Nano and Pro. The model cards include video samples, dataset examples, and code snippets that do not require a local install. This is the fastest way to test Cosmos 3 before committing to hardware. The free tier has rate limits and queues, especially during the first weeks after a major release. Paid inference options are available for larger workloads. You can access the models through the Hugging Face hub. The Hugging Face free inference article explains how free tiers work for open models.
Community inference is useful for validation, not production. You will not get guaranteed latency or privacy. If your robot data is sensitive, you should download weights and run them locally. Still, for running quick benchmarks or comparing Nano and Pro on a small test video, the community path removes setup friction. It also connects you to fine-tuned community variants that appear within days of a release. For a broader view of open-source momentum, see state of open-source on Hugging Face spring 2026.
Key strengths:
- ✅ No local GPU setup required
- ✅ Fast access to model cards and demos
- ✅ Community fine-tunes appear quickly
- ❌ Free tier has queues and rate limits
- ❌ Not suitable for private production data
Who it’s for: Developers who want to evaluate Cosmos 3 before buying GPU hardware.
5. NVIDIA NGC and Isaac SDK Integration , Best for enterprise deployment and production robot stacks
NVIDIA also packages Cosmos 3 through NGC and the Isaac SDK. This is not a separate model but an enterprise access path with container images, optimized kernels, and integration for robot simulation. Teams using Isaac Sim, Isaac Lab, or Omniverse can load Cosmos 3 checkpoints directly into their existing pipeline. The setup supports distributed inference, TensorRT acceleration, and weight streaming. This matters for companies that want to move from a research checkpoint to a deployed robot fleet without rebuilding their stack. You can find the official containers on the NVIDIA developer hub and related code on GitHub.
The NGC path is the most convenient for serious production work, but it ties you to NVIDIA’s software ecosystem. You can still export weights and run them elsewhere under the license terms. Local open-source deployments outside NGC are possible, especially for Nano and Pro. For teams exploring self-hosted LLM workflows, the best open-source LLM models 2026 guide covers hardware and serving options. NGC is not free for all enterprise features, but the model weights themselves remain downloadable.
Key strengths:
- ✅ Optimized containers for Isaac and Omniverse
- ✅ TensorRT acceleration
- ✅ Distributed inference support
- ✅ Direct integration with robot simulation
- ❌ Vendor lock-in risk with NVIDIA ecosystem
- ❌ Enterprise features may require NVIDIA AI Enterprise subscription
Who it’s for: Enterprise robotics teams that already use NVIDIA simulation and deployment tools.
Frequently Asked Questions
What is NVIDIA Cosmos 3?
NVIDIA Cosmos 3 is a family of open-weight physical AI models released on June 17, 2026. It includes 4B, 14B, and 34B parameter models that process video, depth, and robot action data. The models target robotics, autonomous driving simulation, and embodied agent research.
Is NVIDIA Cosmos 3 fully open source?
Nano and Pro use the Apache 2.0 license, which allows commercial use, modification, and redistribution. Ultra uses the NVIDIA Open Model License and restricts at-scale autonomous vehicle fleets. The weights are public, but the licenses are not identical.
Can I run Cosmos 3 on a consumer GPU?
Yes, the 4B Nano model can run on an RTX 4090 or similar hardware. The 14B Pro model requires about 28 GB of VRAM for FP16 inference. The 34B Ultra model is meant for H100 or multi-GPU systems.
What license does Cosmos 3 use?
Cosmos 3 Nano and Pro use Apache 2.0. Cosmos 3 Ultra uses the NVIDIA Open Model License. The Ultra license permits research and most commercial use but adds restrictions for large autonomous vehicle fleets.
How does Cosmos 3 compare to closed physical AI models?
Cosmos 3 gives you inspectable weights and local deployment, unlike closed APIs from OpenAI, Google, or Tesla. Closed models may still lead on proprietary benchmarks. Cosmos 3 is best for teams that need fine-tuning, privacy, or cost control.
Where can I download Cosmos 3 weights?
You can download them from the official NVIDIA developer hub and the public Hugging Face collection. The release includes safetensors weights, config files, and inference scripts. Use NVIDIA NGC for production container images.
What Should You Remember?
- Cosmos 3 is a three-model open-weight physical AI family with 4B, 14B, and 34B parameter options.
- License differences matter: Nano and Pro use Apache 2.0, while Ultra uses the NVIDIA Open Model License.
- Context windows scale from 8,192 tokens to 131,072 tokens for long video and driving scenes.
- Local deployment: Nano runs on a single RTX 4090, Pro needs an 80GB GPU, Ultra needs H100 or DGX class hardware.
- Hugging Face and NVIDIA NGC host the weights and containers for quick testing and production.
- Physical AI open-source progress builds on NVIDIA’s broader open-weight push in 2026.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
