Nemotron 3.5 Lightning 30B A3B VRAM requirements
The smallest GPU memory that runs Nemotron 3.5 Lightning 30B A3B well is 24 GB (Q4_K_M, e.g. RTX Spark). 45 of 132 GPUs run it well. It's a mixture-of-experts model, so all weights must fit in memory but speed depends only on the active experts.
MMLU 81.94 SWE-bench Verified 51.56
VRAM needed per quantization
| Quantization | Weights | 16K context | 32K context | Full 256K |
|---|---|---|---|---|
| Q4_K_M | 23.7 GB | 23.8 GB | 23.9 GB | 25.3 GB |
| Q8_0 | 32.6 GB | 32.7 GB | 32.8 GB | 34.1 GB |
Weights plus fp16 KV cache. Leave ~1 GB headroom for the inference engine.
Run it locally
Each command downloads the GGUF from Hugging Face, then Ollama starts a chat and llama.cpp starts an OpenAI-compatible server with a web UI on port 8080.
ollama run hf.co/bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q4_K_M llama-server -hf bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q4_K_M ollama run hf.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q8_0 llama-server -hf unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q8_0 Cheapest GPUs that can run Nemotron 3.5 Lightning 30B A3B
Lowest launch MSRP among GPUs that run its smallest quantization (Q4_K_M) well, with at least 16K context.
As an Amazon Associate, I earn from qualifying purchases.
GPUs that can run Nemotron 3.5 Lightning 30B A3B
Best quantization that fits on each GPU, with the context and speed it gets. For chips with several memory sizes, the smallest that works is shown.
NVIDIA
- RTX 3090 24 GB Q4_K_M 44K context · ~163 tok/s
- RTX 3090 Ti 24 GB Q4_K_M 44K context · ~170 tok/s
- RTX 4090 24 GB Q4_K_M 44K context · ~170 tok/s
- RTX 4500 Ada 24 GB Q4_K_M 44K context · ~96 tok/s
- RTX A5000 24 GB Q4_K_M 44K context · ~144 tok/s
- RTX PRO 4000 Blackwell 24 GB Q4_K_M 44K context · ~132 tok/s
- RTX Spark 24 GB Q4_K_M 44K context · ~72 tok/s
- Tesla P40 24 GB Q4_K_M 44K context · ~81 tok/s
- RTX 5090 32 GB Q4_K_M 256K context · ~227 tok/s
- RTX 5000 Ada 32 GB Q4_K_M 256K context · ~119 tok/s
- RTX PRO 4500 Blackwell 32 GB Q4_K_M 256K context · ~158 tok/s
- Tesla V100 32GB 32 GB Q4_K_M 256K context · ~159 tok/s
- L40S 48 GB Q8_0 256K context · ~133 tok/s
- RTX 6000 Ada 48 GB Q8_0 256K context · ~143 tok/s
- RTX A6000 48 GB Q8_0 256K context · ~123 tok/s
- RTX PRO 5000 Blackwell 48 GB Q8_0 256K context · ~175 tok/s
- RTX PRO 5000 Blackwell 72GB 72 GB Q8_0 256K context · ~175 tok/s
- RTX PRO 5500 Blackwell 84 GB Q8_0 256K context · ~179 tok/s
- RTX PRO 6000 Blackwell 96 GB Q8_0 256K context · ~203 tok/s
- DGX Spark (GB 10) 128 GB Q8_0 256K context · ~54 tok/s
AMD
- Radeon RX 7900 XTX 24 GB Q4_K_M 44K context · ~165 tok/s
- Radeon AI PRO R9700 32 GB Q4_K_M 256K context · ~128 tok/s
- Ryzen AI Max (Strix Halo) 32 GB Q4_K_M 256K context · ~63 tok/s
- Ryzen AI Max 400 (Gorgon Halo) 64 GB Q8_0 256K context · ~54 tok/s
Intel
- Arc Pro B65 32 GB Q4_K_M 256K context · ~123 tok/s
- Arc Pro B70 32 GB Q4_K_M 256K context · ~123 tok/s
Apple
- M2 24 GB Q4_K_M 44K context · ~27 tok/s
- M3 24 GB Q4_K_M 44K context · ~27 tok/s
- M4 24 GB Q4_K_M 44K context · ~32 tok/s
- M4 Pro 24 GB Q4_K_M 44K context · ~67 tok/s
- M5 24 GB Q4_K_M 44K context · ~40 tok/s
- M5 Pro 24 GB Q4_K_M 44K context · ~73 tok/s
- M6 24 GB Q4_K_M 44K context · ~44 tok/s
- M1 Max 32 GB Q4_K_M 256K context · ~91 tok/s
- M1 Pro 32 GB Q4_K_M 256K context · ~51 tok/s
- M2 Max 32 GB Q4_K_M 256K context · ~91 tok/s
- M2 Pro 32 GB Q4_K_M 256K context · ~51 tok/s
- M3 Max 36 GB Q8_0 256K context · ~59 tok/s
- M3 Pro 36 GB Q8_0 256K context · ~32 tok/s
- M4 Max 36 GB Q8_0 256K context · ~77 tok/s
- M5 Max 36 GB Q8_0 256K context · ~84 tok/s
- M1 Ultra 64 GB Q8_0 256K context · ~98 tok/s
- M2 Ultra 64 GB Q8_0 256K context · ~98 tok/s
- M3 Ultra 96 GB Q8_0 256K context · ~99 tok/s
- M5 Ultra 96 GB Q8_0 256K context · ~131 tok/s