Open Source RoboticsRobots, software & learning

Cosmos-Predict2

Robotics model

Archived NVIDIA Cosmos world foundation model family (0.6B/2B/14B Text2Image and 2B/14B Video2World, plus Bridge action-conditioned and GR00T Dreams samples) with Apache-2.0 code and gated NVIDIA Open Model License weights; superseded by Cosmos-Predict2.5. [1] [2] [4]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Physical AI and world-model researchers; Robotics simulation and synthetic-data engineers; Post-training teams with datacenter NVIDIA GPUs [2] [3] [5]
Parameters
0.6B, 2B, and 14B variants (Text2Image: 0.6B/2B/14B [2] [4]
Embodiments
General-purpose Physical AI visual world simulator (embodiment-agnostic Video2World), GR1 humanoid robot (GR00T… [2] [5]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

Cosmos-Predict2 is NVIDIA's second-generation world foundation model family for Physical AI, spanning Text2Image (0.6B, 2B, 14B) and Video2World (2B, 14B) Diffusion Transformer checkpoints plus robotics post-trained samples for GR00T Dreams (GR1 and DROID) and Bridge action-conditioned video prediction. The GitHub repository (nvidia-cosmos/cosmos-predict2) was archived when checked on 2026-11-21, its reviewed HEAD commit 661da4774b0c is dated 2025-10-07, and its README explicitly advises migrating to nvidia-cosmos/cosmos-predict2.5 for continued support. [1] [2] [4]

Model card summary

Architecture

Diffusion Transformer (DiT) world foundation model family for Text2Image (0.6B, 2B, 14B) and Video2World (2B, 14B) visual world prediction, supporting multi-frame conditioning, optional NATTEN sparse attention for 720p Hopper/Blackwell inference, and action-conditioned Video2World post-training.

[2] [3] [5]

Base model

Cosmos-Predict2 DiT (0.6B / 2B / 14B base checkpoints)

[2]

Parameter count

0.6B, 2B, and 14B variants (Text2Image: 0.6B/2B/14B; Video2World: 2B/14B; Action-Conditioned sample: 2B; GR00T Dreams GR1/DROID samples: 14B)

[2] [4]

Input modalities
rgblanguageproprioception

RGB image/video and text prompts across Video2World/Text2Image; Bridge action-conditioned variant adds end-effector pose/gripper state and 7-DoF action displacement conditioning.

[2] [5]

Output modalities
video-frames

Video2World outputs future RGB video frames (480p and 720p at 10-16 FPS); Text2Image variants output single RGB images.

[2] [3]

Action space

Text/visual prompt conditioning in base Video2World; in the Bridge action-conditioned variant, 7-DoF end-effector displacement (x, y, z, roll, pitch, yaw in gripper frame plus binary open/close gripper command) paired with end-effector state and gripper width, conditioning generated future video frames rather than directly driving hardware.

[2] [5]

Control frequency

10-16 Hz (10 FPS and 16 FPS video generation documented in performance guide)

Refers to generated video frame rate, not real-time closed-loop robot control frequency.

[2] [3]

Trained / supported embodiments
  • General-purpose Physical AI visual world simulator (embodiment-agnostic Video2World)
  • GR1 humanoid robot (GR00T Dreams GR1 post-trained sample)
  • DROID robot manipulation platform (GR00T Dreams DROID post-trained sample)
  • Bridge robot arm dataset embodiment (2B action-conditioned sample)

[2] [5]

Training datasets
  • Pretraining corpus not publicly disclosed or released
  • Bridge dataset (IRASim split) for action-conditioned post-training sample
  • GR00T Dreams GR1 and DROID datasets for post-trained sample checkpoints
  • AgiBotWorld-Alpha (fisheye) and Cosmos-NeMo-Assets documented as post-training case studies

[2] [5]

Interfaces
  • Python API (Video2WorldPipeline / cosmos-predict2 PyPI package)
  • Hugging Face Diffusers (>=0.34.0)
  • PyTorch CLI and torchrun FSDP training scripts

[2] [5]

Checkpoints

CheckpointFormatSizeAccessPinned revision
nvidia/Cosmos-Predict2-0.6B-Text2Imagepytorch4.32 GBgateddd55b6858b22
nvidia/Cosmos-Predict2-2B-Text2Imagesafetensors18.57 GBgatedacdb5fde992a
nvidia/Cosmos-Predict2-14B-Text2Imagesafetensors67.81 GBgated015332720f70
nvidia/Cosmos-Predict2-2B-Video2Worldsafetensors38.14 GBgatedf50c09f5d8ab
nvidia/Cosmos-Predict2-14B-Video2Worldsafetensors210.46 GBgated03b03a377fed
nvidia/Cosmos-Predict2-14B-Sample-GR00T-Dreams-GR1pytorch29.04 GBgated926f34dd8386
nvidia/Cosmos-Predict2-14B-Sample-GR00T-Dreams-DROIDpytorch29.04 GBgatedebc29d30a3fa
nvidia/Cosmos-Predict2-2B-Sample-Action-Conditionedpytorch4.56 GBgated1d2b0a920e76

Sizes are decimal GB sums of sibling files at the pinned HF revision (repos marked safetensors also contain .pt/.pth checkpoints). All 8 repos are gated='auto' under NVIDIA Open Model License.

[2] [4]

Is this for you?

Use cases
  • Video-and-text conditioned future visual world prediction (Video2World)
  • Action-conditioned robotic manipulation video prediction (Bridge post-training)
  • Text-to-image Physical AI scene generation (0.6B, 2B, 14B)
  • Domain-specific world-model post-training (AgiBotWorld-Alpha fisheye, GR00T Dreams GR1/DROID, Cosmos-NeMo-Assets)

[2] [5]

Known limitations
  • Upstream GitHub repository is archived (HEAD 2025-10-07) and explicitly recommends migrating to Cosmos-Predict2.5 for continued support
  • Model weights are gated on Hugging Face under the NVIDIA Open Model License rather than Apache-2.0, and default safety guardrails rely on Llama Guard 3 under its own license
  • High GPU VRAM and storage requirements (26.02-32.54 GB VRAM for 2B models, 48.93-56.38 GB VRAM for 14B models; 14B-Video2World repo is 210.46 GB at HEAD)
  • Pretraining datasets are not released; NATTEN sparse attention speedup is restricted to 720p inference on Hopper/Blackwell GPUs (compute capability 9.0 or 10.0)

[1] [2] [3] [4]

Safety and limitations
  • Outputs simulated images or video frames rather than closed-loop physical robot motor commands; supplied documentation does not establish physical-dynamics accuracy for real robot control.
  • Repository is archived (2025-10-07) and superseded by Cosmos-Predict2.5; users adopting Predict2 weights accept discontinued upstream maintenance.
  • Weights are gated under the NVIDIA Open Model License (not OSI-approved Apache-2.0), and default content moderation uses Llama Guard 3 under its own license.
  • Pretraining datasets are not released; 14B-Text2Image reports OOM on NVIDIA L40S in upstream benchmarks.

[1] [2] [3] [4]

Compute & fine-tuning

Inference compute
  • Minimum VRAM: 26 GB
  • Documented hardware: NVIDIA GB200, B200, H200 SXM/NVL, H100 PCIe/NVL/SXM, H20, L40S, RTX PRO 6000, RTX 6000 Ada, DGX Spark
  • Precision: bf16
  • Reported latency: 25610 ms
  • Edge / consumer capable: No

Required VRAM in performance.md: 26.02 GB (2B-Text2Image), 32.54 GB (2B-Video2World), 48.93 GB (14B-Text2Image), 56.38 GB (14B-Video2World). Latency 25,610 ms (25.61 s) is fastest upstream-reported 480p 16 FPS 2B-Video2World generation on GB200 (79.87 s on H100 PCIe; 14B-Video2World 85.26 s on GB200, 286.46 s on H100 PCIe).

[3]

Fine-tuning

Action-conditioned Bridge post-training example uses torchrun --nproc_per_node=2 with FSDP; exact minimum post-training VRAM is not specified in that guide.

[2] [5]

Host computer

Linux x86-64 with NVIDIA Ampere (RTX 30 Series, A100) or newer GPU; upstream recommends at least 32 GB VRAM for 2B models (26.02-32.54 GB measured) and at least 64 GB VRAM for 14B models (48.93-56.38 GB measured).

[2] [3]

Operating system
  • Linux x86-64 (glibc >=2.31, e.g., Ubuntu >=22.04)

[2]

Dependencies
  • Python 3.10
  • PyTorch 2.6.0 (cu126)
  • CUDA 12.6
  • diffusers>=0.34.0 (optional HF pipeline)
  • NATTEN (optional sparse attention on Hopper/Blackwell)
  • Llama Guard 3 (content input filter)

[2] [3]

Skills
  • Python 3.10 and uv / pip environment management
  • PyTorch and CUDA 12.6 GPU workflows
  • Diffusers or torchrun FSDP post-training

Inferred from documented installation, Diffusers scripts, and torchrun FSDP post-training commands.

[2] [5] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [2] · Model card / documentation [1] [2]

  • Prepare Linux x86-64 (glibc >=2.31) with NVIDIA Ampere or newer GPU, CUDA 12.6 compatible driver, and Python 3.10
  • Create a Python 3.10 virtual environment with uv and install cosmos-predict2[cu126] from PyPI with the cosmos-dependencies extra index
  • Accept NVIDIA Open Model License gating on Hugging Face and download checkpoints or run via diffusers>=0.34.0

[2] [4]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
Cosmos-Predict2 480p 16 FPS Video2World inference latency (documentations/performance.md)2B-Video2World: 25.61 s (GB200), 30.7 s (B200), 50.2 s (H200 SXM), 79.87 s (H100 PCIe), 127.49 s (L40S); 14B-Video2World: 85.26 s (GB200), 176.19 s (H200 SXM), 286.46 s (H100 PCIe), 1036.24 s (L40S).Upstream-reported generation runtime benchmarks at 480p 16 FPS; not independently verified by OSR.
Cosmos-Predict2 720p 16 FPS Video2World NATTEN sparse attention speedup (documentations/performance.md)2B-Video2World: 123.9 s -> 54.0 s (2.3x on B200), 221.7 s -> 89.4 s (2.5x on H200 SXM), 378.5 s -> 149.6 s (2.5x on H100 PCIe); 14B-Video2World: 439.4 s -> 223.1 s (2.0x on B200), 836.9 s -> 412.9 s (2.0x on H200 SXM).Upstream-reported 720p Hopper/Blackwell sparse attention benchmarks; requires compute capability 9.0 or 10.0; not independently verified by OSR.

[3]

Project notes

Capabilities, action conditioning, and robotics post-training

Base Video2World checkpoints generate future visual frames from an initial image or video and a text prompt, with Python (cosmos-predict2 on PyPI) and diffusers>=0.34.0 inference interfaces. For robotics, documentations/post-training_video2world_action.md documents an action-conditioned Bridge recipe (using the IRASim Bridge split) that conditions the 2B Video2World model on end-effector pose, gripper width, and 7-DoF gripper displacement (x, y, z, roll, pitch, yaw plus binary open/close) via 2-GPU torchrun FSDP training; minimum per-GPU VRAM for this post-training recipe is unspecified in the supplied documentation, and the inference memory figures below should not be treated as training requirements. Like other video world models, Cosmos-Predict2 outputs simulated visual frames rather than closed-loop physical robot motor commands, and the reviewed documentation does not establish physical-dynamics fidelity for direct hardware control. [2] [3] [5]

Hardware footprint and upstream-reported performance

Upstream's performance guide specifies Linux x86-64 with NVIDIA Ampere or newer GPUs and CUDA 12.6, with measured VRAM requirements of 26.02 GB (2B-Text2Image), 32.54 GB (2B-Video2World), 48.93 GB (14B-Text2Image), and 56.38 GB (14B-Video2World). Although upstream recommends at least 32 GB VRAM for 2B models and 64 GB for 14B models (14B-Text2Image reports OOM on L40S), the reported 32.54 GB requirement for 2B-Video2World exceeds that general 32 GB recommendation, so the recommendation does not establish that 2B-Video2World fits on a 32 GB GPU. Upstream reports 480p 16 FPS Video2World runtimes ranging from 25.61 s (2B on GB200) and 79.87 s (2B on H100 PCIe) to 85.26 s (14B on GB200) and 286.46 s (14B on H100 PCIe), plus 2.0x–2.6x speedups at 720p using NATTEN sparse attention on Hopper/Blackwell GPUs (compute capability 9.0 or 10.0). Total repository file sizes at the reviewed Hugging Face revisions span 4.32 GB (0.6B-Text2Image) to 210.46 GB (14B-Video2World) across multiple bundled file formats, so the download size for a selected inference configuration may differ. All runtime and VRAM figures are upstream-reported; Open Source Robotics has not run training, fine-tuning, or inference on Cosmos-Predict2. [2] [3] [4]

What we checked

None. Prepared from public upstream sources without hands-on training, fine-tuning, or inference.

Updates

Upstream last updated
Oct 7, 2025 [1]
Latest release
Sep 1, 2025 [4]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
nvidia-cosmos/cosmos-predict2@661da4774b0ca41d082a0ecbeb47550bcf07e03f; HF Cosmos-Predict2-2B-Video2World@f50c09f5d8ab133a90cac3f4886a6471e9ba3f18; HF Cosmos-Predict2-14B-Video2World@03b03a377fede782647afac998f674d9f358e319 · software repository

Separate license matrix

Major Physical AI world foundation model family with Apache-2.0 inference/post-training code and 8 public checkpoints, marked include_partial because weights are gated under the NVIDIA Open Model License, pretraining datasets are unreleased, and the repo is archived in favor of Cosmos-Predict2.5.

LayerAvailability / license
Codeavailable · Apache-2.0
Verified Apache-2.0 in root LICENSE file and README. [1] [2]
Model weightsavailable · NVIDIA Open Model License (gated)
All 8 checkpoints are gated='auto' on Hugging Face under the NVIDIA Open Model License; Llama Guard 3 filter carries its own license. [2] [4]
Training datasetsunavailable · License not fully assessed
Pretraining datasets are not released; post-training tutorials reference external datasets such as Bridge (IRASim), AgiBotWorld-Alpha, and GR00T Dreams. [2] [5]
Documentationavailable · Apache-2.0
[1] [2] [3]

Licensing boundaries and archived status

Licensing splits between code, weights, and auxiliary guardrails. The repository source code is Apache-2.0 (LICENSE), whereas all eight Hugging Face checkpoints are gated (gated: auto) under the NVIDIA Open Model License, and default input content moderation relies on Llama Guard 3 under its own license. Pretraining datasets are not released. Teams evaluating Cosmos-Predict2 should account for both the non-OSI weight license and the archived status of the repository. [1] [2] [4]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Initial 0.8.0-draft.1 model profile prepared from archived GitHub commit 661da4774b0c and 8 verified Hugging Face checkpoints.