Nemotron 3 Nano 4B: VRAM requirements & speed
Nemotron 3 Nano 4B (3.97B dense, nvidia-nemotron-open-model-license). Native context 256k. Smallest GGUF: 2.9 GB.
Memory needed by quantization and context
GGUF file size plus f16 KV cache (and fixed overhead of ~1 GB per GPU not shown). Source: unsloth/NVIDIA-Nemotron-3-Nano-4B-GGUF.
| Quant | File | +8k ctx | +32k ctx | +128k ctx | +256k ctx |
|---|---|---|---|---|---|
| Q4_K_M | 2.9 GB | 3.1 GB | 3.5 GB | 5.1 GB | 7.3 GB |
| Q5_K_M | 3.16 GB | 3.4 GB | 3.8 GB | 5.4 GB | 7.5 GB |
| Q6_K | 4.06 GB | 4.3 GB | 4.7 GB | 6.3 GB | 8.4 GB |
| Q8_0 | 4.23 GB | 4.4 GB | 4.9 GB | 6.5 GB | 8.6 GB |
KV cache size by context length
Computed from the model's attention layout, including a fixed 0.09 GB linear-attention state.
| Context | f16 | q8_0 | q4_0 |
|---|---|---|---|
| 8k | 0.22 GB | 0.16 GB | 0.12 GB |
| 32k | 0.62 GB | 0.37 GB | 0.24 GB |
| 128k | 2.23 GB | 1.23 GB | 0.69 GB |
| 256k | 4.38 GB | 2.37 GB | 1.29 GB |
Which hardware runs Nemotron 3 Nano 4B? · 16k context
| Hardware | Quant | Fit | Memory | Tok/s | Speed source |
|---|---|---|---|---|---|
| RTX 5090 32GB | Q8_0 | fits fully | 4.6 GB | 258 | scaled from measured |
| RTX PRO 6000 96GB | Q8_0 | fits fully | 4.6 GB | 257 | scaled from measured |
| Mac Studio M5 Ultra 256GB | Q8_0 | fits fully | 4.6 GB | 176 | estimated |
| Mac Studio M5 Ultra 512GB (ships late Oct 2026) | Q8_0 | fits fully | 4.6 GB | 176 | estimated |
| RTX 4090 24GB | Q8_0 | fits fully | 4.6 GB | 167 | scaled from measured |
| RTX 5080 16GB | Q8_0 | fits fully | 4.6 GB | 162 | scaled from measured |
| RTX 5070 Ti 16GB | Q8_0 | fits fully | 4.6 GB | 152 | scaled from measured |
| 2x RTX 3090 (48GB total) | Q8_0 | fits fully | 4.6 GB | 151 | estimated |
| RTX 3090 24GB | Q8_0 | fits fully | 4.6 GB | 146 | scaled from measured |
| AMD RX 7900 XTX 24GB | Q8_0 | fits fully | 4.6 GB | 130 | scaled from measured |
| RTX 5070 12GB | Q8_0 | fits fully | 4.6 GB | 124 | scaled from measured |
| Mac Studio M5 Max 128GB | Q8_0 | fits fully | 4.6 GB | 108 | scaled from measured |
| MacBook Pro 14/16 M5 Max 128GB | Q8_0 | fits fully | 4.6 GB | 108 | scaled from measured |
| Mac Studio (2025) M3 Ultra 96GB | Q8_0 | fits fully | 4.6 GB | 103 | scaled from measured |
| RTX 4070 12GB | Q8_0 | fits fully | 4.6 GB | 89 | scaled from measured |
| RTX 5060 Ti 16GB | Q8_0 | fits fully | 4.6 GB | 84 | scaled from measured |
| RTX 3060 12GB (32GB RAM) | Q8_0 | fits fully | 4.6 GB | 66 | scaled from measured |
| MacBook Pro 14/16 (2024) M4 Max 128GB | Q8_0 | fits fully | 4.6 GB | 61 | scaled from measured |
| 16GB VRAM class (RTX 4060 Ti 16GB, 32GB RAM) | Q8_0 | fits fully | 4.6 GB | 57 | scaled from measured |
| 8GB VRAM class (RTX 4060 8GB, 32GB RAM) | Q8_0 | fits fully | 4.6 GB | 54 | scaled from measured |
| NVIDIA DGX Spark 128GB | Q8_0 | fits fully | 4.6 GB | 53 | scaled from measured |
| AMD Ryzen AI Max+ PRO 495 (Ryzen AI Halo) 192GB | Q8_0 | fits fully | 4.6 GB | 52 | estimated |
| MacBook Pro 14/16 M5 Pro 48GB | Q8_0 | fits fully | 4.6 GB | 51 | scaled from measured |
| AMD Ryzen AI Max+ 395 (Strix Halo) 128GB | Q8_0 | fits fully | 4.6 GB | 46 | scaled from measured |
| Mac mini (2024) M4 Pro 64GB | Q8_0 | fits fully | 4.6 GB | 37 | scaled from measured |
| Mac mini M6 32GB | Q8_0 | fits fully | 4.6 GB | 29 | scaled from measured |
| MacBook Air M5 32GB | Q8_0 | fits fully | 4.6 GB | 23 | scaled from measured |
| Mac mini / iMac M4 16GB | Q8_0 | fits fully | 4.6 GB | 23 | scaled from measured |
Discrete GPUs are modelled with 32 GB of system RAM available for offload. Fits (at least with offload) on every listed configuration.
MLX conversions for Apple Silicon
No MLX conversion yet: we found no verified mlx-community 4-bit or 8-bit repo for Nemotron 3 Nano 4B. Use the GGUF files with llama.cpp or ollama on a Mac. How we check.
Links
Model card · GGUF files · Open in the advisor · last verified 2026-10-20
FAQ
How much VRAM does Nemotron 3 Nano 4B need?
The smallest quant (Q4_K_M) is 2.9 GB, so you need roughly 4 GB for an 8k context including KV cache and overhead. At 256k context the f16 KV cache alone is 4.4 GB.
Can I run Nemotron 3 Nano 4B on a RTX 5090 32GB?
Yes: on RTX 5090 32GB it fits fully as Q8_0 at about 258 tok/s (scaled from measured). 28 of 28 hardware configurations fit it fully at 16k context.
What is the maximum context for Nemotron 3 Nano 4B?
256k tokens natively. Memory, not the model, is usually the limit: see the KV cache table above.
All models: model index · by memory: 8 GB · 12 GB · 16 GB · 24 GB · 32 GB · 48 GB · 64 GB · 96 GB · 128 GB · 192 GB · 256 GB · 512 GB. Numbers use the same fit math as the advisor; sources and methods.