All models › nvidia
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16
nvidia · on Arena since 12 Oct 2026
Arena score—Not voted on Arena yet
Size124.5 B12 B active per token (mixture of experts)
4-bit download75 GBweights only, Q4_K_M
Max context1024Ktokens, from the model config
Where to use it for free
- Free API on OpenRouter: model id
nvidia/nemotron-3.5-super-vl:free. Try it side by side · OpenRouter page - 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
New: waiting for Arena votes
Released on 12 Oct 2026. Arena has not voted on this model yet. As soon as it does, the model moves into the ranking automatically, usually the next night. Until then you can still check its size, memory needs and speed below, and try it for free.
Speed on popular hardware
| Hardware | GPU memory | Fits? | Speed |
|---|---|---|---|
| RTX 3060 12 GB | 12 GB | Does not fit | — |
| RTX 4060 Ti 16 GB | 16 GB | Does not fit | — |
| RTX 3090 | 24 GB | Does not fit | — |
| RTX 4090 | 24 GB | Does not fit | — |
| RTX 5090 | 32 GB | GPU + 64 GB RAM | 17 tok/s |
| RX 7900 XTX | 24 GB | Does not fit | — |
| 2 × RTX 3090 | 48 GB | GPU + 64 GB RAM | 22 tok/s |
| 4 × RTX 3090 | 96 GB | Fits in GPU | 80 tok/s |
| Mac M4 Max 128 GB | 96 GB | Fits in GPU | 51 tok/s |
| Ryzen AI Max+ 395 128 GB | 96 GB | Fits in GPU | 24 tok/s |
| 8 × H100 SXM | 640 GB | Fits in GPU | 257 tok/s |
Estimate at 4-bit with 32K context, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 70% ÷ data read per token; real numbers vary by about ±30%.
GPU memory needed
| Level | 8K context | 32K context | 128K context | Quality vs full |
|---|---|---|---|---|
| Q3_K_M | 63 GB | 63 GB | 64 GB | +4.7% noticeable loss |
| Q4_K_M | 76 GB | 76 GB | 77 GB | +1.2% very small loss |
| Q5_K_M | 89 GB | 89 GB | 90 GB | +0.4% near-lossless |
| Q6_K | 103 GB | 103 GB | 104 GB | +0.2% near-lossless |
| Q8_0 | 133 GB | 133 GB | 134 GB | ≈ 0% same as full |
| F16 | 250 GB | 250 GB | 251 GB | ≈ 0% same as full |
Weights + context cache (from the model's architecture) + runtime. Quality: increase in perplexity against the full model, from llama.cpp tests, scaled by model size. Simulate your own setup.