Nanbeige4.3 33B VRAM requirements: which GPUs can run it locally?
33.12B params 512K max context Tool useReasoning
The smallest GPU memory that runs Nanbeige4.3 33B well is 24 GB (Q4_K_M, e.g. RTX Spark). 45 of 132 GPUs run it well and 2 more are a tight fit.
VRAM needed per quantization
| Quantization | Weights | 16K context | 32K context | Full 512K |
|---|---|---|---|---|
| Q4_K_M | 19.4 GB | 22.5 GB | 25.7 GB | 120.7 GB |
| Q8_0 | 32.8 GB | 35.9 GB | 39.1 GB | 134.2 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
ollama run hf.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF:Q4_K_M llama.cpp
llama-server -hf bartowski/Nanbeige_Nanbeige4.3-33B-GGUF:Q4_K_M Ollama
ollama run hf.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF:Q8_0 llama.cpp
llama-server -hf bartowski/Nanbeige_Nanbeige4.3-33B-GGUF:Q8_0 Cheapest GPUs that can run Nanbeige4.3 33B
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 Nanbeige4.3 33B
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 4000 Ada 20 GB Q4_K_M 3K context · ~17 tok/s tight fit
- RTX 3090 24 GB Q4_K_M 23K context · ~43 tok/s
- RTX 3090 Ti 24 GB Q4_K_M 23K context · ~46 tok/s
- RTX 4090 24 GB Q4_K_M 23K context · ~46 tok/s
- RTX 4500 Ada 24 GB Q4_K_M 23K context · ~20 tok/s
- RTX A5000 24 GB Q4_K_M 23K context · ~35 tok/s
- RTX PRO 4000 Blackwell 24 GB Q4_K_M 23K context · ~31 tok/s
- RTX Spark 24 GB Q4_K_M 23K context · ~14 tok/s
- Tesla P40 24 GB Q4_K_M 23K context · ~16 tok/s
- RTX 5090 32 GB Q4_K_M 63K context · ~82 tok/s
- RTX 5000 Ada 32 GB Q4_K_M 63K context · ~26 tok/s
- RTX PRO 4500 Blackwell 32 GB Q4_K_M 63K context · ~41 tok/s
- Tesla V100 32GB 32 GB Q4_K_M 63K context · ~41 tok/s
- L40S 48 GB Q8_0 76K context · ~24 tok/s
- RTX 6000 Ada 48 GB Q8_0 76K context · ~27 tok/s
- RTX A6000 48 GB Q8_0 76K context · ~21 tok/s
- RTX PRO 5000 Blackwell 48 GB Q8_0 76K context · ~37 tok/s
- RTX PRO 5000 Blackwell 72GB 72 GB Q8_0 198K context · ~37 tok/s
- RTX PRO 5500 Blackwell 84 GB Q8_0 258K context · ~39 tok/s
- RTX PRO 6000 Blackwell 96 GB Q8_0 319K context · ~50 tok/s
- DGX Spark (GB 10) 128 GB Q8_0 480K context · ~8 tok/s
AMD
- Radeon RX 7900 XT 20 GB Q4_K_M 3K context · ~37 tok/s tight fit
- Radeon RX 7900 XTX 24 GB Q4_K_M 23K context · ~44 tok/s
- Radeon AI PRO R9700 32 GB Q4_K_M 63K context · ~29 tok/s
- Ryzen AI Max (Strix Halo) 32 GB Q4_K_M 63K context · ~12 tok/s
- Ryzen AI Max 400 (Gorgon Halo) 64 GB Q8_0 157K context · ~8 tok/s
Intel
- Arc Pro B65 32 GB Q4_K_M 63K context · ~28 tok/s
- Arc Pro B70 32 GB Q4_K_M 63K context · ~28 tok/s
Apple
- M2 24 GB Q4_K_M 23K context · ~5 tok/s
- M3 24 GB Q4_K_M 23K context · ~5 tok/s
- M4 24 GB Q4_K_M 23K context · ~6 tok/s
- M4 Pro 24 GB Q4_K_M 23K context · ~13 tok/s
- M5 24 GB Q4_K_M 23K context · ~7 tok/s
- M5 Pro 24 GB Q4_K_M 23K context · ~14 tok/s
- M6 24 GB Q4_K_M 23K context · ~8 tok/s
- M1 Max 32 GB Q4_K_M 63K context · ~18 tok/s
- M1 Pro 32 GB Q4_K_M 63K context · ~9 tok/s
- M2 Max 32 GB Q4_K_M 63K context · ~18 tok/s
- M2 Pro 32 GB Q4_K_M 63K context · ~9 tok/s
- M3 Max 36 GB Q8_0 16K context · ~8 tok/s
- M3 Pro 36 GB Q8_0 16K context · ~4 tok/s
- M4 Max 36 GB Q8_0 16K context · ~11 tok/s
- M5 Max 36 GB Q8_0 16K context · ~13 tok/s
- M1 Ultra 64 GB Q8_0 157K context · ~15 tok/s
- M2 Ultra 64 GB Q8_0 157K context · ~15 tok/s
- M3 Ultra 96 GB Q8_0 319K context · ~16 tok/s
- M5 Ultra 96 GB Q8_0 319K context · ~23 tok/s