Can You Run Llama 4 Maverick on an RTX 5090? Practical Limits
The previous guide incorrectly treated Llama 4 Maverick like a 70B model that fits entirely in 15–18GB of VRAM. Meta's model card says Maverick is a mixture-of-experts model with 17B activated parameters and 400B total parameters. Activated parameters describe computation per token; they do not mean that only 17B weights must be stored.
This page replaces an older version that contained outdated or unsupported claims.
Verified facts
Meta lists Llama 4 Maverick as 17B activated parameters, 128 experts and 400B total parameters with a 1M context window.
Meta distributes BF16 and FP8 weights and says the official FP8 weights fit on a single H100 DGX host, not a 32GB consumer GPU.
A third-party quantization may use system RAM, CPU or disk offloading, but that is different from keeping the complete model in RTX 5090 VRAM.
The old Ollama commands, 15–18GB claim and 80–120 token-per-second benchmark were not supported by reproducible evidence and have been removed.
A realistic local deployment decision
- Use a smaller model if you need predictable single-GPU performance on 32GB VRAM.
- If experimenting with community quantizations, verify the exact file size, runtime support, RAM requirement, context length and measured speed before downloading.
- Treat benchmark results as configuration-specific: quantization, offload split, prompt length and memory bandwidth all change performance.