Written by Jakub Rusinowski · Last updated July 21, 2026
Model library → Cosmos 3 → Cosmos 3 Super
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.
| 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 |
| Specified at | BF16 |
| System RAM | 256 GB |
| Record updated | 2026-07-21 |
OpenMDW-1.1 — commercial use permitted. Commercial use permitted. No usage restrictions beyond attribution.
Modelled on a reference NVIDIA RTX 4090 (24 GB), with no KV cache (this record has no published architecture). Speed figures are ESTIMATES from the memory-bandwidth roofline described on the methodology page, not benchmarks we ran — rows marked measured come from published or reader-submitted runs. VRAM here includes the KV cache, so it reads higher than the headline figure above, which does not.
| Quant | Weights | VRAM needed | Est. speed | Fit on 24 GB |
|---|---|---|---|---|
| Q2_K | 21.0 GB | 21.8 GB | ~57 tok/s (est.) | Tight fit |
| Q3_K_M | 27.3 GB | 28.1 GB | ~7 tok/s (est.) | Offloads to system RAM (slow) |
| Q4_K_M | 38.6 GB | 39.4 GB | ~5 tok/s (est.) | Offloads to system RAM (slow) |
| Q5_K_M | 45.4 GB | 46.2 GB | ~4 tok/s (est.) | Offloads to system RAM (slow) |
| Q6_K | 52.5 GB | 53.3 GB | ~4 tok/s (est.) | Offloads to system RAM (slow) |
| Q8_0 | 68.0 GB | 68.8 GB | — | Won't fit |
| F16 | 128.0 GB | 128.8 GB | — | Won't fit |
Want the memory numbers alone, at every quantization level and your own context length? Use the Cosmos 3 Super VRAM calculator.
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.
The cheapest catalogued GPU that runs Cosmos 3 Super is the Apple M5 Pro (64 GB).
Install Ollama, then run:
ollama run cosmos-3Weights on Hugging Face: nvidia/Cosmos3-Super.
Quality scores as published by the model's authors or an independent evaluator — not throughput, and not measured by us.
| 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) |
Best for: robotics, autonomous vehicles, world simulation, synthetic data, post training.
← All Cosmos 3 models | VRAM calculator | Check your own hardware