Llama 4 Scout 17B-16E Instruct: how much VRAM does it need?

Llama 4 Scout 17B-16E Instruct has 108.64B parameters in 48 layers. It uses grouped-query attention: 40 query heads share 8 key-value heads, so its KV cache is 5× smaller than full multi-head attention would need. It is a Mixture-of-Experts model with about 17.00B of its 108.64B parameters active per token: memory has to hold all of them, while each token passes through only the active share. It takes up to 10,485,760 tokens of context; one user at that length needs 1920.0 GiB of KV cache at FP16 on top of the weights. Its repository is gated on Hugging Face, so LLMScale ships its architecture in a bundled database.

Llama 4 Scout 17B-16E Instruct memory by precision and context

One user, FP16 KV cache and vLLM's overhead on one GPU at its defaults for a card under 70 GB. Quantizing shrinks the weights, not the KV cache, so the saving narrows as the context grows.

Which single GPUs hold it

• BF16 (205.4 GiB): AMD MI325X, B300 (Blackwell Ultra) 270GB, AMD MI355X, Apple M3 Ultra (512GB)
• FP8 (106.2 GiB): NVIDIA DGX Spark (GB10), H200 141GB, Apple M2 Ultra, B200 180GB
• INT4 (59.64 GiB): RTX PRO 5000 Blackwell (72GB), A100 80GB, H100 SXM 80GB, H100 PCIe 80GB

At 8k context for one user, the smallest cards first. A discrete card is budgeted at vLLM's default 92% of its memory, a unified-memory device at the share its GPU can address, and each carries vLLM's overhead at its own defaults.

How many users fit

Users at 8k context each, on one card. ✗: the weights leave no room for even one.

Fine-tuning Llama 4 Scout 17B-16E Instruct

Sequence length 2,048, batch size 1, gradient checkpointing on, Adam. Full fine-tuning keeps 16 bytes per parameter before activations; QLoRA stores the frozen base in 4-bit NF4.

Where the numbers come from

Weights are parameters × bytes per parameter. The KV cache is 2 × layers × key-value heads × head dimension × bytes × context × users. For Llama 4 Scout 17B-16E Instruct that is 192.0 KiB per token at FP16. These are capacity-planning figures: the serving engine — vLLM, TGI, llama.cpp — and its paged-attention efficiency still move the real number.

More on KV cache size and on GQA, MQA and MLA.

Work with Llama 4 Scout 17B-16E Instruct

• Size it for your workload
• Fine-tuning memory
• vLLM parameters
• Model anatomy
• Its config.json, explained
• What its files say about serving it

Other Llama models

• Llama 3.2 1B Instruct
• Llama 3.2 3B Instruct
• Llama 3.1 8B Instruct
• Llama 3.3 70B Instruct
• Llama 3.1 70B Instruct
• Llama 3.1 405B Instruct