---
title: "Cosmos Predict 2.5"
description: "NVIDIA world foundation model: text/image/video in, world video out. Native 2B/14B recipes; gated checkpoints. 2B text-to-world is the verified route."
modelId: "cosmos-predict-2.5"
pageSource: authored
---
Cosmos Predict 2.5 is NVIDIA's **world foundation model** for physical AI: text, image, or video in, world video out. The paper is [Cosmos Predict 2.5](https://arxiv.org/abs/2511.00062) (see also [research.nvidia.com/labs/dir/cosmos-predict2.5](https://research.nvidia.com/labs/dir/cosmos-predict2.5)). This catalog entry assembles the model from WorldFoundry's canonical denoiser, conditioner, scheduler, decoder, and loader roles **without Diffusers**.
Two sizes are recorded: **2B** and **14B**. Structural checks in this repository match the official EMA transformer **569/569** (2B) and **729/729** (14B) tensors, Reason1 language weights **338/338**, and the shared **Wan VAE** **194/194**. Image/video conditioning and positive classifier-free guidance go through those shared runner contracts, not a model-private backend. NVIDIA access is **gated**; the recorded license is `other:nvidia-open-model-license`.
The verified WorldFoundry recipe is the **2B text-to-world** route at **1280×704, 93 frames, 35 steps, 16 fps**, flow-UniPC scheduling. The 14B route still needs end-to-end GPU output validation. Aliases `cosmos-predict2p5` and `cosmos-predict-2p5` resolve here.
## What it is for
Predict future world video from text, a first frame, or a video prefix. Variants are `cosmos-predict-2.5-2b` and `cosmos-predict-2.5-14b`. Default artifact: `cosmos-predict2.5-2b.mp4` (14B writes `cosmos-predict2.5-14b.mp4`). Score world-generation quality on [WorldModelBench](/docs/evaluation/benchmark-hub/worldmodelbench) and physical plausibility on [Physics-IQ](/docs/evaluation/benchmark-hub/physics-iq). NVIDIA has since pointed new work at Cosmos 3; this page stays on the 2.5 card that is actually wired.
## Runtime contract
`CosmosPredict2p5Pipeline` unifies text/image/video-to-world in one **rectified-flow** model with a **Reason1** language conditioner and a shared Wan VAE. Prompt conditioning, positive guidance, image/video latent conditioning, and decoding all run through the shared native diffusion roles. Upstream NVIDIA cookbooks also mention robot / action-conditioned and policy variants (RoboCasa, LIBERO); those are not extra WorldFoundry variants on this card.
## Running in WorldFoundry
The pipeline invokes the native recipes directly in `worldfoundry-unified-cu128`. Launch `worldfoundry-eval run cosmos-predict-2.5-2b` with `--pipeline.task-profile text-to-world-validation`. Stage `nvidia/Cosmos-Predict2.5-2B` (and 14B if you are testing that line) under NVIDIA's access rules. Structural tensor match is not the same as GPU output parity — 14B still pending.
The in-tree network is an inference-only PyTorch `MinimalV1LVGDiT`. It uses WorldFoundry attention and timestep primitives. NVIDIA EMA conversion and the `Denoiser` adapter live in `worldfoundry/base_models/diffusion_model/models/denoisers/cosmos2p5.py`, not in the network package. Architecture reference: `nvidia-cosmos/cosmos-predict2.5` at `a2c298b0a3df3778b973fe65e9e58877b292d8a7`.
## Minimal eval
The 2B text-to-world route is verified. Import [WorldModelBench](/docs/evaluation/benchmark-hub/worldmodelbench) official-shaped results and read `world_model_average`. `official-run` needs `WORLDFOUNDRY_WORLDMODELBENCH_JUDGE`; full judge execution is not verified.