Cosmos-Transfer2.5

NVIDIA, United States 路 October 2025

World model 路 Open weights, NVIDIA Open Model 路 Also written Cosmos Transfer 2.5, Cosmos-Transfer2.5-2B

Parameters
2.4 B
2,358,047,744 parameters according to the Hugging Face model card.
GPU memory
65.4 GB
reported by the authors
Inference latency
not published
per action step
Weights
NVIDIA Open Model
NVIDIA Open Model License for the checkpoint, code Apache-2.0.

What it is

Cosmos-Transfer2.5 is NVIDIA's control-net style world model for sim to real and real to real translation. From a text prompt and up to four control videos (Canny edges, blurred RGB, segmentation masks, depth maps) it renders a new photoreal video that keeps the layout and motion of the input, so one simulator rollout or one real demonstration can become many visually different training samples. It is built on Cosmos-Predict2.5-2B and is 3.5 times smaller than Cosmos-Transfer1 according to the paper. In NVIDIA's own robot test, a diffusion policy trained on 100 real demonstrations plus 500 Transfer2.5 variants succeeded in 24 of 30 trials across a base scene and nine changed scenes, against 5 of 30 with standard image augmentation. It is heavy to run, with 65.4 GB of GPU memory on the model card.

Architecture

Backbone
Cosmos-Predict2.5-2B diffusion transformer with added control branches for up to four control videos.
Action head
None. It changes the appearance of a video while keeping its structure, which is what makes it usable for augmenting robot demonstrations.
Parameters
2,358,047,744 parameters according to the Hugging Face model card.
Pretraining data
Inherits the Cosmos-Predict2.5-2B base, which was trained on 200M curated video clips.
Embodiments
Semi-humanoid with two Kinova Gen3 arms (policy experiment in the paper)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 2.4 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp328.8 GBTraining master weights, rarely used for inference
bf16 or fp164.4 GBThe usual way these checkpoints are served
int82.2 GBQuantised, expect some loss of precision on fine motions
int41.1 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference memory, reported65.4 GB source

The model card lists 65.4 GB of GPU memory and single GPU generation times with segmentation control of 285.83 s on a B200, 719.4 s on an H100 NVL, 870.3 s on an H100 PCIe and 2,326.6 s on an H20. Only BF16 is tested. Frame counts that are multiples of 93 work best.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Real robot, two Kinova Gen3 arms (Cosmos paper protocol) bimanual apple and bowl pick-and-place, base setting plus nine new scenarios, 3 trials each, UNet diffusion policy trained with Transfer2.5 augmented demonstrations
    80%
    success rateAuthors' own evaluation, 24 of 30 trials. The same policy with standard image augmentation reached 5 of 30, without augmentation 1 of 30. The number measures a policy trained on augmented data, not the world model alone.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Real robot test in the paper: a semi-humanoid platform with two 7-DoF Kinova Gen3 arms and a head camera, a bimanual apple and bowl pick-and-place task, 100 teleoperated demonstrations and five augmented variants of each. A UNet diffusion policy trained with the variants succeeded in 24 of 30 trials (base setting plus nine new scenarios, three trials each), against 5 of 30 with standard image augmentation and 1 of 30 without augmentation.

Fine tuning it yourself

Used as a data augmentation tool rather than fine tuned: generate variants of existing demonstrations from edge, blur, segmentation or depth controls, then train the policy on originals and variants. Code is Apache-2.0 in the cosmos-transfer2.5 repository.

Where it helps, where it does not

Strengths

  • Large measured effect on policy robustness in the authors' test: 24 of 30 successes against 5 of 30 with standard augmentation.
  • Keeps scene layout and motion while changing appearance, driven by edge, blur, segmentation or depth controls.
  • 3.5 times smaller than Cosmos-Transfer1, according to the paper.
  • Open weights under the NVIDIA Open Model License.

Limits

  • Needs 65.4 GB of GPU memory and about 12 minutes per clip on an H100 NVL (719.4 s).
  • The robot evidence is one task on one platform with 30 trials.
  • Does not produce actions or predict the outcome of actions; it only re-renders video.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-11-12.

Try a policy on a real arm

A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.