What GPU do you need to run Nemotron Cascade 2 30B A3B?
30B params 256K max context ReasoningTool use
The smallest GPU memory that runs Nemotron Cascade 2 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 79.8 SWE-bench Verified 50.2
VRAM needed per quantization
| Quantization | Weights | 16K context | 32K context | Full 256K |
|---|---|---|---|---|
| Q4_K_M | 23.0 GB | 23.1 GB | 23.2 GB | 24.6 GB |
| Q8_0 | 31.3 GB | 31.4 GB | 31.5 GB | 32.8 GB |
Weights plus fp16 KV cache. Leave ~1 GB headroom for the inference engine.
Cheapest GPUs that can run Nemotron Cascade 2 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 Cascade 2 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 161K context · ~284 tok/s
- RTX 3090 Ti 24 GB Q4_K_M 161K context · ~305 tok/s
- RTX 4090 24 GB Q4_K_M 161K context · ~305 tok/s
- RTX 4500 Ada 24 GB Q4_K_M 161K context · ~131 tok/s
- RTX A5000 24 GB Q4_K_M 161K context · ~233 tok/s
- RTX PRO 4000 Blackwell 24 GB Q4_K_M 161K context · ~204 tok/s
- RTX Spark 24 GB Q4_K_M 161K context · ~91 tok/s
- Tesla P40 24 GB Q4_K_M 161K context · ~105 tok/s
- RTX 5090 32 GB Q8_0 119K context · ~434 tok/s
- RTX 5000 Ada 32 GB Q8_0 119K context · ~139 tok/s
- RTX PRO 4500 Blackwell 32 GB Q8_0 119K context · ~217 tok/s
- Tesla V100 32GB 32 GB Q8_0 119K context · ~218 tok/s
- L40S 48 GB Q8_0 256K context · ~209 tok/s
- RTX 6000 Ada 48 GB Q8_0 256K context · ~232 tok/s
- RTX A6000 48 GB Q8_0 256K context · ~186 tok/s
- RTX PRO 5000 Blackwell 48 GB Q8_0 256K context · ~325 tok/s
- RTX PRO 5000 Blackwell 72GB 72 GB Q8_0 256K context · ~325 tok/s
- RTX PRO 5500 Blackwell 84 GB Q8_0 256K context · ~338 tok/s
- RTX PRO 6000 Blackwell 96 GB Q8_0 256K context · ~434 tok/s
- DGX Spark (GB 10) 128 GB Q8_0 256K context · ~66 tok/s
AMD
- Radeon RX 7900 XTX 24 GB Q4_K_M 161K context · ~291 tok/s
- Radeon AI PRO R9700 32 GB Q8_0 119K context · ~155 tok/s
- Ryzen AI Max (Strix Halo) 32 GB Q8_0 119K context · ~62 tok/s
- Ryzen AI Max 400 (Gorgon Halo) 64 GB Q8_0 256K context · ~66 tok/s
Intel
- Arc Pro B65 32 GB Q8_0 119K context · ~147 tok/s
- Arc Pro B70 32 GB Q8_0 119K context · ~147 tok/s
Apple
- M2 24 GB Q4_K_M 161K context · ~30 tok/s
- M3 24 GB Q4_K_M 161K context · ~30 tok/s
- M4 24 GB Q4_K_M 161K context · ~36 tok/s
- M4 Pro 24 GB Q4_K_M 161K context · ~83 tok/s
- M5 24 GB Q4_K_M 161K context · ~46 tok/s
- M5 Pro 24 GB Q4_K_M 161K context · ~93 tok/s
- M6 24 GB Q4_K_M 161K context · ~52 tok/s
- M1 Max 32 GB Q8_0 119K context · ~97 tok/s
- M1 Pro 32 GB Q8_0 119K context · ~48 tok/s
- M2 Max 32 GB Q8_0 119K context · ~97 tok/s
- M2 Pro 32 GB Q8_0 119K context · ~48 tok/s
- M3 Max 36 GB Q8_0 256K context · ~73 tok/s
- M3 Pro 36 GB Q8_0 256K context · ~36 tok/s
- M4 Max 36 GB Q8_0 256K context · ~99 tok/s
- M5 Max 36 GB Q8_0 256K context · ~111 tok/s
- M1 Ultra 64 GB Q8_0 256K context · ~194 tok/s
- M2 Ultra 64 GB Q8_0 256K context · ~194 tok/s
- M3 Ultra 96 GB Q8_0 256K context · ~198 tok/s
- M5 Ultra 96 GB Q8_0 256K context · ~291 tok/s