Llama 4 Maverick VRAM requirements: which GPUs can run it?
The smallest GPU memory that runs Llama 4 Maverick well is 256 GB (Q4_K_M, e.g. M3 Ultra). 2 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 85.5
VRAM needed per quantization
| Quantization | Weights | 16K context | 32K context | Full 128K |
|---|---|---|---|---|
| Q4_K_M | 226.1 GB | 229.0 GB | 231.9 GB | 249.5 GB |
| Q8_0 | 396.6 GB | 399.5 GB | 402.4 GB | 420.0 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. Both fetch the vision projector (mmproj) automatically.
llama-server -hf unsloth/Llama-4-Maverick-17B-128E-Instruct-GGUF:Q4_K_M Split into 5 files, which Ollama can't pull from Hugging Face yet; llama.cpp downloads the whole set.
llama-server -hf unsloth/Llama-4-Maverick-17B-128E-Instruct-GGUF:Q8_0 Split into 9 files, which Ollama can't pull from Hugging Face yet; llama.cpp downloads the whole set.
GPUs that can run Llama 4 Maverick
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.