Primary use stated or inferred cautiously from official documentation.
N
NVIDIA / nvidia/Cosmos3-H-Embodied-Nano
Cosmos3-H-Embodied-Nano
Research checkpoint published on October 20, 2026, for surgical robotics. A single model offers three modes: direct action-to-video dynamics, inverse video-to-action dynamics, and a policy conditioned on image, text, and history. It derives from Cosmos3-Nano, with a Qwen3-VL reasoning tower and fine-tuning of the generative tower. It is not three independent models nor an authorized medical device.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | nvidia/Cosmos3-H-Embodied-Nano |
|---|---|
| Model type | Robotics · Video · Multimodal |
| Access model | Open weights |
| License | NVIDIA Open Model License Agreement |
| Deployment | Self-hosted |
| Release | 20 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image · Video · Actions |
| Exit | Video · Actions |
| Context window | 2.4-second clips; up to 16 previous actions |
| maximum output | 25 frames and 24 action steps at 10 Hz, depending on the mode |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · PyTorch |
Research in simulation, imitation, and policy evaluation with Open-H data. The protocol uses 2.4-second clips, one context frame, and 24 future frames, with 24 action steps at 10 Hz. Bimanual actions are normalized by dataset and padded with a mask: they should not be interpreted as universal commands across robots.
Its license is the NVIDIA Open Model License Agreement, not OpenMDW 1.1. The output is adjusted to a 256p-class resolution, with dimensions according to the aspect ratio. The HTTP server is an implementation for self-deployment, not a validated commercial API. Drift grows in long sequences; do not use the checkpoint for clinical operation without independent controls, validation, and authorization.
03 / CAPABILITIES
What it can do
Orientation
Research in simulation, imitation, and policy evaluation with Open-H data. The protocol uses 2.4-second clips, one context frame, and 24 future frames, with 24 action steps at 10 Hz. Bimanual actions are normalized by dataset and padded with a mask: they should not be interpreted as universal commands across robots.
Context
2.4-second clips; up to 16 previous actions · 25 frames and 24 action steps at 10 Hz, depending on the mode
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Input modalities
TextImageVideoActions
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
Public 16B parameter checkpoint for self-hosted infrastructure. We have not verified a commercial endpoint or token rates: numerical costs are left blank. Downloadable weights do not make inference free. Budget for GPU, memory, video encoding, and operation; review the license of each variant and its dependencies. Playback frames per second do not equate to generation performance.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
The provider evaluates 640 clips from ten public families with reserved episodes. Direct dynamics: PSNR 26.1; inverse: normalized action MSE 0.009. The policy records an average MSE of 0.052 and a best-of-four of 0.033: these are different selections. These metrics do not measure clinical safety nor guarantee instrument tracking. They have not been reproduced by Inferama.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| Cosmos3-H held-out 640 forward PSNR ↑ | 26,1 dB | Score | View source ↗ |
| Cosmos3-H held-out 640 inverse action MSE ↓ | 0,009 | Score | View source ↗ |
| Cosmos3-H held-out 640 policy mean action MSE ↓ | 0,052 | Score | View source ↗ |
| Cosmos3-H held-out 640 policy best-of-4 action MSE ↓ | 0,033 | Score | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Cosmos3-H-Embodied-Nano
Version added to Inferama's verified catalog.
07 / SOURCES