unsloth
Nemotron 3 Nano 30B A3B
Nemotron 3 Nano 30B A3B is a 31.6B mixture-of-experts nemotron_h_moe model from unsloth, with about 3.0B parameters active per token. Configured context length is 1,048,576 tokens. License: other. Quant sizes are the GGUF file bytes published on Hugging Face when a file exists.
Parameters31.6B / 3.0B active
Context1024K
Attentiongqa
Licenseother
52 layers · 2 KV heads · head dim 128 · GGUF repo · config
Quant file sizes
| Quant | File size | Source |
|---|---|---|
| BF16 | 58.8 GB | Hugging Face file bytes |
| UD-Q8_K_XL | 37.7 GB | Hugging Face file bytes |
| Q8_0 | 31.3 GB | Hugging Face file bytes |
| UD-Q6_K_XL | 31.2 GB | Hugging Face file bytes |
| Q6_K | 31.2 GB | Hugging Face file bytes |
| UD-Q5_K_XL | 25.6 GB | Hugging Face file bytes |
| Q5_K_M | 24.4 GB | Hugging Face file bytes |
| Q5_K_S | 22.3 GB | Hugging Face file bytes |
| UD-Q4_K_XL | 21.3 GB | Hugging Face file bytes |
| Q4_K_M | 22.9 GB | Hugging Face file bytes |
| Q4_K_S | 20.5 GB | Hugging Face file bytes |
| IQ4_NL | 16.9 GB | Hugging Face file bytes |
| IQ4_XS | 16.9 GB | Hugging Face file bytes |
| Q4_1 | 18.7 GB | Hugging Face file bytes |
| Q4_0 | 17.0 GB | Hugging Face file bytes |
| UD-Q3_K_XL | 18.6 GB | Hugging Face file bytes |
| Q3_K_M | 18.6 GB | Hugging Face file bytes |
| Q3_K_S | 16.9 GB | Hugging Face file bytes |
| UD-IQ3_XXS | 16.9 GB | Hugging Face file bytes |
| UD-Q2_K_XL | 18.6 GB | Hugging Face file bytes |
| Q2_K | 16.9 GB | Hugging Face file bytes |
| UD-IQ2_XXS | 16.9 GB | Hugging Face file bytes |
On common machines
| Device | Verdict | Quant | Speed |
|---|---|---|---|
| GeForce RTX 4060 Ti 8GB | Expert offload | Q4_K_M | 57–78 |
| GeForce RTX 3060 12GB | Expert offload | Q4_K_M | 71–97 |
| GeForce RTX 4060 Ti 16GB | Expert offload | Q4_K_M | 57–78 |
| GeForce RTX 4090 | Tight fit | Q4_K_S | 219–299 |
| GeForce RTX 5090 | Tight fit | UD-Q5_K_XL | 323–441 |
| Apple M4 Pro | Tight fit | IQ4_NL | 77–105 |
| Apple M5 | Won't fit | Q2_K | — |
| CPU only · DDR5-5600 | Runs on CPU | Q5_K_S | 18–25 |
Smallest devices that fit it fully
- RTX 4000 Ada — Tight fit, IQ4_NL, 92–125 tok/s
- Radeon RX 7900 XT — Tight fit, IQ4_NL, 203–277 tok/s
- GeForce RTX 4090 — Tight fit, Q4_K_S, 219–299 tok/s
- GeForce RTX 3090 Ti — Tight fit, Q4_K_S, 219–299 tok/s
- GeForce RTX 3090 — Tight fit, Q4_K_S, 204–278 tok/s
- NVIDIA L4 — Tight fit, UD-Q4_K_XL, 63–86 tok/s
Run it
Commands for the GeForce RTX 4060 Ti 16GB estimate (Q4_K_M).
llama.cpp
# Expert offload is approximate. llama.cpp can pin experts to CPU: # -ot "\\.ffn_.*_exps\\.=CPU" llama-server -hf unsloth/Nemotron-3-Nano-30B-A3B-GGUF:Q4_K_M -c 8192 -ngl 99
Ollama
ollama run nemotron
LM Studio
lms get unsloth/Nemotron-3-Nano-30B-A3B-GGUF # Then load quant Q4_K_M