NVIDIA64B~39 GB VRAM at BF16

Cosmos 3 Super — VRAM, speed & local setup

Name: LLM Configurator — GPU VRAM Checker
Author: LLM Configurator

Written by Jakub Rusinowski · Last updated July 21, 2026

The largest, highest-fidelity tier — a 64B-parameter two-tower MoT built on a dense 32B transformer backbone. Super is the model you post-train: NVIDIA positions it for building custom robotics and autonomous-vehicle world models, generating high-quality synthetic training video and predicting long-horizon action sequences with strong physics accuracy. NVIDIA reports Cosmos 3 Super takes the highest overall PAI-Bench-G scores on both text-to-video and image-to-video across all models tested, including closed-source. Datacenter-class hardware: ~128 GB VRAM at BF16 (roughly halved at FP8), so a Hopper/Blackwell multi-GPU node (H100/H200/B200 class). Omnimodal I/O (text, image, video, audio, action). OpenMDW-1.1. Specs from launch coverage and the Hugging Face model card — verify before relying on them.

Cosmos 3 Super needs about 39 GB of VRAM at BF16 — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.

Logic90

Creative92

Coding35

VRAM and speed by quantization

Quoted against NVIDIA RTX 4090 (24 GB). Weights plus overhead only: this model's architecture is unpublished, so no KV cache is added.

Quant	VRAM	Speed (est.)	Fit
Q2_K 2.63 bpw	21.8 GB	~57 tok/s	Tight
Q3_K_M 3.41 bpw	28.1 GB	~7 tok/s	Offload
Q4_K_M 4.83 bpw	39.4 GB	~5 tok/s	Offload
Q5_K_M 5.67 bpw	46.2 GB	~4 tok/s	Offload
Q6_K 6.56 bpw	53.3 GB	~4 tok/s	Offload
Q8_0 8.50 bpw	68.8 GB	—	Too big
F16 16.00 bpw	128.8 GB	—	Too big

Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. Cosmos 3 Super VRAM calculator →

Get Cosmos 3 Super running

The cheapest catalogued GPU that runs Cosmos 3 Super is the NVIDIA DGX Spark (128 GB).

Buy this hardware NVIDIA DGX Spark (128GB) — 128 GB VRAM · 150 W board power Deploy in the cloud now NVIDIA A40 on RunPod — from $0.44/hr · rate checked 2026-08

or compare on Vast.ai

As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.

Affiliate disclosure: Some links on this page are affiliate links — if you buy through them, LLM Configurator may earn a commission at no extra cost to you. As an Amazon Associate, LLM Configurator earns from qualifying purchases.

NVIDIA DGX Spark (128GB)

128 GB VRAM · 150 W board power

2026 prices are volatile — check the current listing.

Check price on Amazon

How to run Cosmos 3 Super

Install Ollama, then run:

$ollama run cosmos-3

Weights on Hugging Face: nvidia/Cosmos3-Super ↗

Specifications

Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong.

Parameters: 64B total (32B dense transformer backbone)
Context window: Omnimodal I/O (video/action)
Architecture: Two-tower Mixture-of-Transformers — autoregressive reasoner (VLM) + diffusion generator, shared multimodal attention
Provider: NVIDIA
Licence: OpenMDW-1.1
Practical deployment: Workstation (large unified memory or pro GPU)
Specified at: BF16
System RAM: 256 GB
Record updated: 2026-07-21

LicenceOpenMDW-1.1Commercial use permitted

Commercial use permitted. No usage restrictions beyond attribution.

Quality and use cases

Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.

Best forroboticsautonomous vehiclesworld simulationsynthetic datapost training

Benchmark	Score	Provenance
PAI-Bench-G Image-to-Video (Overall)	0.81 / 1	vendor-claimed · NVIDIA Cosmos 3 (launch)
PAI-Bench-G Text-to-Video (Overall)	0.77 / 1	vendor-claimed · NVIDIA Cosmos 3 (launch)
R-Bench / Physics-IQ / RoboLab	1 / 1 open-source SOTA	vendor-claimed · NVIDIA Cosmos 3 (launch)

Can I run Cosmos 3 Super on my GPU?

Other Cosmos 3 sizes

Cosmos 3 Super — frequently asked questions

How much VRAM does Cosmos 3 Super need?

About 39 GB at BF16 — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.

Does Cosmos 3 Super run on an RTX 4090 (24 GB)?

No. Cosmos 3 Super needs about 39 GB at BF16, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.

How do I run Cosmos 3 Super locally?

Install Ollama and run `ollama run cosmos-3`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.

What other sizes does Cosmos 3 come in?

Cosmos 3 Super (39 GB), Cosmos 3 Nano (10 GB), Cosmos 3 Edge (3 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.