N
NVIDIA / nvidia/Cosmos3-Nano-Sim-Auto

Cosmos3-Nano-Sim-Auto

Distilled causal variant of Cosmos3-Nano-Transfer-Auto, published by NVIDIA on October 20, 2026, for closed-loop driving simulation. It produces synchronized multiview video and LiDAR from controls and text. It supports up to eleven cameras. The ability to respond to controls during a sequence distinguishes it from the bidirectional generator, but it does not constitute an autonomous driving policy.

01 / SUMMARY

The essential

ORIENTATIONCausal multiview simulation

Primary use stated or inferred cautiously from official documentation.

CONTEXTText, WSM controls, and optional visual reference

Maximum input capacity when published by the source.

EXITUp to 11 cameras; 480p or 720p; playback from 10 to 30 fps

Documented maximum generation limit.

ACCESSHugging Face · downloadable checkpoint

Verified availability channels.

MODEL TYPERobotics · Video · Multimodal

Primary functional family and verified specialties.

LICENSEOpenMDW 1.1

Declared terms for API access, model weights, or self-hosting.

02 / TECHNICAL SHEET

Limits and integration

API IDnvidia/Cosmos3-Nano-Sim-Auto
Model typeRobotics · Video · Multimodal
Access modelOpen weights
LicenseOpenMDW 1.1
DeploymentSelf-hosted
Release20 OCT 2026
Knowledge cutoffNot published
EntranceText · Image · Video · LiDAR
ExitVideo · LiDAR
Context windowText, WSM controls, and optional visual reference
maximum outputUp to 11 cameras; 480p or 720p; playback from 10 to 30 fps
ReasoningNot published
Published tools
Structured OutingsMultiview video and LiDAR; aligned LiDAR controls are required to generate LiDAR.
Batch processingNot published
Prompt cacheNot published
Fine-tuningNot published
Verified platformsHugging Face · PyTorch
BEST SUITED FOR

Interactive simulation with standard and fisheye cameras at 480p or 720p. The declared playback is between 10 and 30 fps; actual inference speed depends on GPU, views, and steps. We do not present the 297 frames and seven cameras from the benchmark as a general product limit nor inherit Transfer-Auto's maximum.

WORTH MONITORING

Sim-Auto uses guidance 1 in the recorded results, compared to guidance 3 for Transfer-Auto. Comparisons with Transfer2.5 also use different prompts. Do not mix those configurations nor interpret a single FID value as simulator safety. Review drift, sensor synchronization, and out-of-distribution scenarios before evaluating a driving policy.

03 / CAPABILITIES

What it can do

01

Orientation

Interactive simulation with standard and fisheye cameras at 480p or 720p. The declared playback is between 10 and 30 fps; actual inference speed depends on GPU, views, and steps. We do not present the 297 frames and seven cameras from the benchmark as a general product limit nor inherit Transfer-Auto's maximum.

02

Context

Text, WSM controls, and optional visual reference · Up to 11 cameras; 480p or 720p; playback from 10 to 30 fps

03

Tools and integration

Not published

04

Access

Hugging Face · downloadable checkpoint

Input modalities
TextImageVideoLiDAR
04 / PRICES

Documented cost

ConceptWorthUnit/condition
Standard inputNo published official API rateStandard API rate
Cached inputNot publishedReading reused prefixes
Cache write or storageNot publishedThe condition varies by provider
Standard outputSelf-hosted infrastructure costMay include reasoning tokens
Batch inputNot publishedAsynchronous processing
Batch outputNot publishedAsynchronous 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

“

NVIDIA publishes results in Driving Simulation Evaluation with RGB and LiDAR quality separately. We preserve FID and FVD-I3D for T2V and I2V with guidance 1. The values correspond to this causal variant and improve as they decrease; Transfer-Auto's values are not copied. Inferama's own evaluation and validation in real driving are pending.

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-Nano-Sim-Auto

Version added to Inferama's verified catalog.

07 / SOURCES

Traceability