NVIDIA Cosmos hardware requirements: VRAM per Cosmos3 model and the GPU servers for it

Cosmos hardware requirements at a glance

NVIDIA Cosmos is a family of world foundation models that generate, predict and reason about video of physical scenes for robotics, driving and factory work; the current models are Cosmos3 Edge, Nano and Super with 4, 16 and 64 billion parameters. For the Nano model, NVIDIA's NIM support matrix asks for 58 GiB of GPU memory per card to generate video in BF16 without offloading, which one NVIDIA RTX PRO 6000 Blackwell Server Edition with 96 GB provides. Super in FP8 needs 93 GiB per card and fits an H200 NVL with 141 GB with room to spare.

The figures in this guide come from the support matrix of NVIDIA NIM for Cosmos, last updated on 21 October 2026, the Hugging Face model cards of the Cosmos3 models and NVIDIA's Cosmos documentation, all read on 25 October 2026. A NIM is NVIDIA's packaged inference container: it starts a model with a profile chosen for the GPU it finds. The memory floors apply to the NIM; the open runtimes listed on the model cards, vLLM-Omni, SGLang and Hugging Face Diffusers, publish no floor of their own.

Two rules from the support matrix decide the hardware before any table. The GPUs in one deployment must be identical, and "Do not add VRAM across devices": each GPU must meet the floor on its own, in total memory and in free memory at start-up. Eight cards of 96 GB therefore count as eight separate 96 GB devices for the Super Generator in BF16, each of which has to hold a full copy of the profile's working set.

Our guides to Isaac Sim and Isaac Lab servers and Omniverse system requirements cover simulation and digital twins, and our guide to video generation GPU requirements covers creative video models; this one covers the Cosmos models that generate, transform and interpret the video those pipelines produce.

The Cosmos models in October 2026

Cosmos3 is a "Mixture-of-Transformers" model, as its model card describes it: an autoregressive transformer for discrete tokens such as text and a diffusion transformer for continuous outputs such as video, trained together. The NIM runs it in two modes from one container image, nvcr.io/nim/nvidia/cosmos3:2.1.0. The Generator handles text-to-video, image-to-video, video-to-video, action prediction and Transfer through /v1/infer. The Reasoner answers questions about images and video through an OpenAI-compatible chat API. Each container runs one mode.

Hugging Face model cards and "Files" pages of nvidia/Cosmos3-Nano, Cosmos3-Super, Cosmos3-Edge and Cosmos-Reason2-8B; NVIDIA Cosmos Transfer2.5 model matrix; read 25 October 2026. Transformer size is the sum of the safetensors files in the transformer folder.

The Nano repository holds 35 GB in total, of which the diffusion transformer is 30.3 GB: 16 billion parameters at two bytes each, as BF16 stores them, plus the text, vision and audio parts. The Super transformer alone is 128 GB, so in BF16 it does not fit on any single PCIe card. The Cosmos3 model cards state that "Only BF16 precision is tested"; the FP8 and NVFP4 profiles appear only in the NIM's support matrix.

The earlier families are still documented. Cosmos-Predict2.5-2B generates video from text, an image or a video; Cosmos-Transfer2.5-2B, in NVIDIA's model matrix, needs one GPU with 65.4 GB for inference; Cosmos-Reason2 needs at least 24 GB for the 2B and 32 GB for the 8B model, according to NVIDIA's prerequisites page. Multiview inference of Predict2.5 for driving, with seven camera views, "requires a minimum of 8 GPUs with at least 80GB memory each".

VRAM per NIM profile: Generator, Transfer and Reasoner

The support matrix lists one floor per profile, in GiB per GPU, and a minimum of system memory per host. Offloading moves parts of the model to system memory while it runs: "model" offload moves whole components, "layer" offload moves the transformer layer by layer and needs the least GPU memory.

NVIDIA NIM for Cosmos support matrix, last updated 21 October 2026, read 25 October 2026. Ranges depend on whether the guardrail models stay on the GPU or are offloaded. Offload rows are single-GPU rows; rows without offload allow 1, 2, 4 or 8 GPUs.

Transfer is the mode that turns a source video into a new one under control inputs, such as depth, edges or segmentation from a simulator, which is how simulated scenes are made to look like camera footage. It peaks higher than plain generation, and the NIM checks it separately at start-up: a deployment can serve generation and reject Transfer. NVIDIA's override, NIM_ALLOW_UNSAFE_TRANSFER=1, is "for diagnosis only" and can end in out-of-memory failures, so size for the Transfer row if your pipeline uses it.

NVIDIA NIM for Cosmos support matrix, 21 October 2026, read 25 October 2026. On discrete GPUs the Reasoner measures free memory after a 4 GiB runtime reserve.

Precision also has a hardware floor. FP8 profiles need compute capability 8.9, which the L40S, the H200 NVL (9.0) and all RTX PRO Blackwell cards (12.0) meet. The Reasoner's NVFP4 profile needs 10.0, so it runs on Blackwell cards and not on the H200 NVL.

Which GPU meets which floor

We compare each floor with the memory the driver reports: 95.6 GiB on an RTX PRO 6000 Server Edition and 140.4 GiB on an H200 NVL. These are not our usual LLM rule of 90 % less 3 GiB, because the NIM checks the floors itself at start-up. The margins are our arithmetic.

Floors from NVIDIA's NIM for Cosmos support matrix (21 October 2026); visible memory from NVIDIA driver figures as used on our site (RTX PRO 4500 Server Edition derived from our usable-memory figure of 25.7 GiB); assignment is our reading, read 25 October 2026.

The RTX PRO 6000 Server Edition row needs two remarks. The Super Generator in FP8 has a floor of 93 GiB, 2.6 GiB below what the card shows, and the NIM also checks free memory at start-up; for "H100, RTX PRO 6000 Blackwell, and lower-throughput discrete GPUs", NVIDIA's own advice is "Default to Nano, especially for video". With model offload, Super in FP8 drops to 64 GiB per card and 80 GiB of system memory, which fits comfortably at the cost of speed. For Transfer with Super, the card needs model offload (76 GiB) or layer offload.

The H200 NVL clears the Super Generator and Super Transfer in FP8 with room to spare, and NVIDIA's guidance for H200- and B200-class GPUs is to use Super when the workload needs it. It does not reach the Super Generator in BF16 (150 GiB), which no PCIe card does. Its single-card Super Reasoner in BF16 leaves 5.4 GiB above the 135 GiB floor, and about 1.4 GiB once the 4 GiB reserve is taken, so we would run the FP8 profile there.

For the smaller cards, a 24 GB card shows less than 24 GiB to the driver, so it clears the Nano Reasoner's 23.1 GiB floor by under 1 GiB at best, before the 4 GiB reserve; by our reading these cards run the Edge Reasoner. The L40S, with 48 GB, leaves little room above the 44 GiB floor of the Nano Generator in FP8 without offload, too little to plan on, and runs it with model offload (38 GiB) or layer offload. The RTX PRO 4500 Server Edition with 32 GB runs the Nano Reasoner. Its Generator floors with layer offload (31 GiB) leave under 1 GiB, which we would not plan on. NVIDIA's support matrix does not mention MIG or vGPU, so we size Cosmos on whole GPUs.

We supply the RTX PRO 6000 Server Edition and the H200 NVL in servers built to order with two, four or eight cards. Send us the Cosmos modes you plan to run and the clip volume per day, and our team replies with a configuration.

Prerequisites: system memory, disk and software

System memory is listed per host and grows with offloading. Nano without offload needs 40 GiB of host RAM; the offload rows for Super need 80 GiB in FP8 and 150 GiB in BF16. On a server that runs one Generator container per GPU, the host RAM adds up. Four Nano containers need 160 GiB by our multiplication, which our entry configuration of 384 GB holds. Four Super containers with BF16 layer offload would need 600 GiB, which only the 768 GB of our four-card configuration covers. Our article on how much RAM a GPU server needs explains why we size host memory at twice the GPU memory.

For the Cosmos3 NIM, NVIDIA's prerequisites page recommends driver 595.58.03 or newer and names its tested setup (Ubuntu 24.04.5 LTS, Docker 27.0.3, Container Toolkit 1.20.1) without setting minimums for disk or software versions. For the earlier Predict2.5 and Transfer2.5 NIMs, the prerequisites page of the previous documentation version (3.0.0) lists Linux on x86_64 or Arm with glibc 2.35 or later, an NVIDIA driver from 580, Docker 23.0.1 or later, NVIDIA Container Toolkit 1.16.2 or later, at least 100 GB of disk for container and model, and at least 90 GB of RAM. The Transfer2.5 NIM requires compute capability 9.0 or higher, which excludes Ada cards such as the L40S; its release 1.1.0 added the RTX PRO 6000 Workstation and Server Editions. For disk, plan the model cache in addition: the Cosmos3-Nano repository is 35 GB, the Super transformer 128 GB, and each further NIM image and fine-tuned checkpoint adds to that.

Without the NIM, the Cosmos3 model cards list PyTorch, vLLM-Omni, Diffusers and SGLang. The Nano model card names vLLM 0.21.0 for CUDA 13 drivers or 0.19.1 for CUDA 12.8 with a vllm-cosmos3 add-on package, and recommends --enable-layerwise-offload on GPUs with less memory. Containers on a multi-GPU server need each GPU assigned explicitly; our guide to Docker GPU assignment without Kubernetes shows how.

What more GPUs buy: latency and throughput

More cards help in two ways, and the NIM has a profile for each. The throughput profile runs one engine replica per GPU, so a four-card server generates four clips at once. The latency profile spreads one request over the cards, with guidance parallelism 2 and sequence parallelism of half the GPU count, so one clip finishes sooner. The Super Generator rows without offload allow 1, 2, 4 or 8 GPUs, but each GPU still has to meet the 93 GiB (FP8) or 150 GiB (BF16) floor.

The model cards give a sense of the times. On one RTX PRO 6000 Blackwell Server Edition, Cosmos3-Edge generated a 480p image-to-video clip of 189 frames in 36.29 seconds with vLLM-Omni and 38.98 seconds with PyTorch, at batch size 1; an H100 NVL took 35.60 seconds. The Super model card states that a 50-step video takes about 55 seconds on eight H200 GPUs and about three minutes on two. These are NVIDIA's figures with NVIDIA's settings, not results we reproduced.

For the Reasoner, the Edge model card lists one RTX PRO 6000 Blackwell Server Edition serving 64 concurrent short requests (its workload 50/100/1) at 10.27 requests and 1,026 output tokens per second. Reasoning over video is a vision-language task, so one card serves many users at once, while video generation keeps a card busy for each clip.

Licences: weights, container and production use

The Cosmos3 weights are released under OpenMDW-1.1. Its grant is to "deal in the Model Materials without restriction"; it names no territory or field of use, imposes no restrictions on outputs and asks you to keep the licence and notices when you distribute the materials. Rights end if you sue claiming the materials infringe a patent or copyright. The model cards say the models are "ready for commercial and non-commercial use".

Cosmos-Reason2 is under the NVIDIA Open Model License of 24 October 2025. It allows commercial use worldwide, asks for the notice "Licensed by NVIDIA Corporation under the NVIDIA Open Model License" when you distribute the model and a "Built on NVIDIA Cosmos" credit when you distribute a product that uses it or train another model with it, and ends your rights automatically if you bypass or disable a safety guardrail without an equivalent one. Neither licence excludes the EU.

The NIM container is a separate matter: it is governed by the NVIDIA Software License Agreement and NVIDIA's product-specific terms. NVIDIA's NIM FAQ, as quoted in our guide to which NVIDIA software is free, states that using NIM in production requires an NVIDIA AI Enterprise licence per GPU. The H200 NVL includes five years of AI Enterprise. With the open runtimes from the model cards, the model licence alone applies. Your legal counsel confirms how these terms fit your project.

Server configurations for Cosmos synthetic data

For a team that generates synthetic training video and runs a video reasoning service, three configurations cover most cases. Power figures are our estimates from the card and processor ratings, confirmed by a measurement on the delivered system.

Configurations from our AI server page; floors from NVIDIA's support matrix; power by our estimate (cards, processors and 15 % for the rest), 25 October 2026.

The RTX PRO 6000 also brings what the H200 NVL lacks for the neighbouring work: RT cores for Isaac Sim, which does not support GPUs without them, as our robotics server guide explains, and NVFP4 for the Reasoner. A site that simulates in Isaac Sim, augments the clips with Cosmos Transfer and trains a perception model on the result often ends up with RTX PRO 6000 Server Edition cards for both, and adds H200 NVL nodes only when Super in FP8 becomes the default. Our comparison of four RTX PRO 6000 Server Edition and four H200 NVL sets out that choice for language models.

We deliver these AI servers assembled and tested before delivery, on one EU contract and invoice, with the operating system, driver, CUDA and container runtime installed on request. Describe your Cosmos pipeline, from simulator to training job, and we size the cards and host memory to it.

FAQ