Llama 3.3 70B VRAM requirements: which GPUs can run it locally?
The smallest GPU memory that runs Llama 3.3 70B well is 48 GB (Q4_K_M, e.g. RTX Spark). 22 of 132 GPUs run it well.
MMLU 86 SWE-bench Verified 24.8
VRAM needed per quantization
| Quantization | Weights | 16K context | 32K context | Full 128K |
|---|---|---|---|---|
| Q4_K_M | 39.6 GB | 44.5 GB | 49.4 GB | 78.6 GB |
| Q8_0 | 69.8 GB | 74.7 GB | 79.6 GB | 108.9 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/unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_M llama-server -hf unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_M llama-server -hf unsloth/Llama-3.3-70B-Instruct-GGUF:Q8_0 Split into 2 files, which Ollama can't pull from Hugging Face yet; llama.cpp downloads the whole set.
Cheapest GPUs that can run Llama 3.3 70B
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 Llama 3.3 70B
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
- L40S 48 GB Q4_K_M 27K context · ~20 tok/s
- RTX 6000 Ada 48 GB Q4_K_M 27K context · ~22 tok/s
- RTX A6000 48 GB Q4_K_M 27K context · ~18 tok/s
- RTX PRO 5000 Blackwell 48 GB Q4_K_M 27K context · ~31 tok/s
- RTX Spark 48 GB Q4_K_M 27K context · ~7 tok/s
- RTX PRO 5000 Blackwell 72GB 72 GB Q4_K_M 106K context · ~31 tok/s
- RTX PRO 5500 Blackwell 84 GB Q8_0 46K context · ~18 tok/s
- RTX PRO 6000 Blackwell 96 GB Q8_0 85K context · ~24 tok/s
- DGX Spark (GB 10) 128 GB Q8_0 128K context · ~4 tok/s
AMD
- Ryzen AI Max (Strix Halo) 64 GB Q4_K_M 80K context · ~6 tok/s
- Ryzen AI Max 400 (Gorgon Halo) 64 GB Q4_K_M 80K context · ~6 tok/s
Apple
- M3 Max 48 GB Q4_K_M 27K context · ~9 tok/s
- M4 Max 48 GB Q4_K_M 27K context · ~13 tok/s
- M4 Pro 48 GB Q4_K_M 27K context · ~6 tok/s
- M5 Max 48 GB Q4_K_M 27K context · ~14 tok/s
- M5 Pro 48 GB Q4_K_M 27K context · ~7 tok/s
- M1 Max 64 GB Q4_K_M 80K context · ~9 tok/s
- M1 Ultra 64 GB Q4_K_M 80K context · ~13 tok/s
- M2 Max 64 GB Q4_K_M 80K context · ~9 tok/s
- M2 Ultra 64 GB Q4_K_M 80K context · ~13 tok/s
- M3 Ultra 96 GB Q8_0 85K context · ~8 tok/s
- M5 Ultra 96 GB Q8_0 85K context · ~11 tok/s