All models › nvidia
NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB
nvidia · on Arena since 7 Oct 2026
Arena score—Not voted on Arena yet
Size22.4 B9 B active per token (mixture of experts)
4-bit download14 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
New: waiting for Arena votes
Released on 7 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.
Speed on popular hardware
| Hardware | GPU memory | Fits? | Speed |
|---|---|---|---|
| RTX 3060 12 GB | 12 GB | GPU + 64 GB RAM | 32 tok/s |
| RTX 4060 Ti 16 GB | 16 GB | Fits in GPU | 35 tok/s |
| RTX 3090 | 24 GB | Fits in GPU | 114 tok/s |
| RTX 4090 | 24 GB | Fits in GPU | 123 tok/s |
| RTX 5090 | 32 GB | Fits in GPU | 219 tok/s |
| RX 7900 XTX | 24 GB | Fits in GPU | 117 tok/s |
| 2 × RTX 3090 | 48 GB | Fits in GPU | 111 tok/s |
| 4 × RTX 3090 | 96 GB | Fits in GPU | 105 tok/s |
| Mac M4 Max 128 GB | 96 GB | Fits in GPU | 67 tok/s |
| Ryzen AI Max+ 395 128 GB | 96 GB | Fits in GPU | 31 tok/s |
| 8 × H100 SXM | 640 GB | Fits in GPU | 339 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 | 12 GB | 12 GB | 13 GB | +7.8% noticeable loss |
| Q4_K_M | 14 GB | 14 GB | 15 GB | +2.1% small loss |
| Q5_K_M | 17 GB | 17 GB | 18 GB | +0.7% very small loss |
| Q6_K | 19 GB | 19 GB | 20 GB | +0.3% near-lossless |
| Q8_0 | 24 GB | 25 GB | 25 GB | ≈ 0% same as full |
| F16 | 45 GB | 46 GB | 46 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.