How much VRAM does Llama 4 Scout (109B) need?

About Llama 4 Scout (109B)

How much VRAM does an Llama 4 Scout (109B) need? At bf16 — the reference precision, no quality loss — a reference 8K-context, batch-1 deployment needs 223.7GB, which is enough to tell you the smallest fitting configuration is 1× AMD MI325X 256GB. That number comes straight out of our GPU calculator's architecture-aware model, the same path the interactive calculator uses, not a size-based estimate. Llama 4 Scout (109B) is a Llama 4 model built from 109B total parameters, with only 17.0B active per token via mixture-of-experts routing, and most of that VRAM at batch size 1 is model weights rather than KV cache or runtime overhead — which is exactly why quantization moves the needle so much. Switching to fp8 cuts the weight footprint roughly in half; AWQ (weight-only int4) goes further, at some quality cost on harder tasks; and NVFP4 lands in a similar range to AWQ but with native throughput on Blackwell GPUs instead of an emulated dtype. The table above walks through all four options with the exact VRAM total and the smallest GPU configuration each one unlocks. Context length adds its own pressure on top of the weights: every additional token of context grows the KV cache, so a 128K-context deployment of this model needs meaningfully more headroom than the 8K reference case shown as the headline number — the context-sensitivity table breaks that out at 32K and 128K wherever this model's native window allows it.