Written by Jakub Rusinowski · Last updated July 21, 2026
Model library → Cosmos 3 → Cosmos 3 Nano
The mid-tier workhorse — a 16B-parameter two-tower MoT on a dense 8B backbone, tuned for high-quality video and action reasoning in a fraction of a second. Nano is the balance point: much cheaper to serve than Super while still leading open models (NVIDIA reports it matches the second-best result on R-Bench). Weights are ~32 GB at BF16, but NVIDIA recommends an RTX PRO 6000 / 96 GB-class GPU for full-context BF16 inference because video generation activations are memory-hungry. 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 Nano needs about 10 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 | 16B total (8B 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 | 96 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 | 5.3 GB | 6.1 GB | ~150 tok/s (est.) | Fits comfortably |
| Q3_K_M | 6.8 GB | 7.6 GB | ~130 tok/s (est.) | Fits comfortably |
| Q4_K_M | 9.7 GB | 10.5 GB | ~104 tok/s (est.) | Fits comfortably |
| Q5_K_M | 11.3 GB | 12.1 GB | ~93 tok/s (est.) | Fits comfortably |
| Q6_K | 13.1 GB | 13.9 GB | ~84 tok/s (est.) | Fits comfortably |
| Q8_0 | 17.0 GB | 17.8 GB | ~69 tok/s (est.) | Fits comfortably |
| F16 | 32.0 GB | 32.8 GB | ~6 tok/s (est.) | Offloads to system RAM (slow) |
Want the memory numbers alone, at every quantization level and your own context length? Use the Cosmos 3 Nano VRAM calculator.
or compare on Vast.ai from $0.35/hr (typical low · varies)
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 Nano is the Intel Arc B570 (10 GB).
Install Ollama, then run:
ollama run cosmos-3Weights on Hugging Face: nvidia/Cosmos3-Nano.
Quality scores as published by the model's authors or an independent evaluator — not throughput, and not measured by us.
| Benchmark | Score | Provenance |
|---|---|---|
| R-Bench (Image-to-Video, open-source) | 2 / 2 2nd best (rank) | vendor-claimed · NVIDIA Cosmos 3 (launch) |
Best for: robotics, world simulation, video generation, action reasoning, synthetic data.
← All Cosmos 3 models | VRAM calculator | Check your own hardware