All models › nvidia
Arena score1152#33 of 82 in Image, paid models included
Size64.6 Bparameters
4-bit download39 GBweights only, Q4_K_M
Max context256Ktokens, from the model config
Where to use it for free
- The lab's own chat: build.nvidia.com (free with daily limits)
- Download and run it at home: official weights · GGUF files for Ollama and LM Studio
Arena scores
| Arena | Score | Rank | Votes |
|---|---|---|---|
| Image | 1152 | #33 of 82 | 18K |
Rank among all models, paid ones included. Source: Arena leaderboard dataset.
Does it fit on popular hardware?
| Hardware | GPU memory | Fits? |
|---|---|---|
| RTX 3060 12 GB | 12 GB | GPU + 64 GB RAM |
| RTX 4060 Ti 16 GB | 16 GB | GPU + 64 GB RAM |
| RTX 3090 | 24 GB | GPU + 64 GB RAM |
| RTX 4090 | 24 GB | GPU + 64 GB RAM |
| RTX 5090 | 32 GB | GPU + 64 GB RAM |
| RX 7900 XTX | 24 GB | GPU + 64 GB RAM |
| 2 × RTX 3090 | 48 GB | Fits in GPU |
| 4 × RTX 3090 | 96 GB | Fits in GPU |
| Mac M4 Max 128 GB | 96 GB | Fits in GPU |
| Ryzen AI Max+ 395 128 GB | 96 GB | Fits in GPU |
| 8 × H100 SXM | 640 GB | Fits in GPU |
Estimate at 4-bit, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 70% ÷ data read per token; real numbers vary by about ±30%.
GPU memory needed
| Level | Total |
|---|---|
| Q3_K_M | 36 GB |
| Q4_K_M | 43 GB |
| Q5_K_M | 50 GB |
| Q6_K | 57 GB |
| Q8_0 | 72 GB |
| F16 | 133 GB |
Weights + text encoders and decoder (about 3 GB) + runtime. Simulate your own setup.
Image score over time
12001400
MayJunJulAugSeptOct
Cosmos3-Super-Text2Image 1152Best paid model
Arena score from 10 Apr 2026 to 7 Oct 2026.
Similar free models in Image
- hunyuan-image-3.0 1150 · 83 B
- flux-2-dev 1146 · 32.2 B
- Cosmos3-Super-Text2Image (Agentic) 1165
- qwen-image-2512 1125 · 20.4 B
- hidream-o1-image 1116 · 8.8 B
- krea-2-turbo 1113 · 12.8 B