Llama 4 Scout 17B 16E Instruct (Q2_K) — 639.4 GBon NVIDIA H100 640GB

Q2_K NVIDIA H100 640GB

Overview

Llama 4 Scout 17B 16E Instruct is a 17B parameter moe language model by Meta, with code, multilingual, tool-calls, vision capabilities. It supports a context window of up to 10,485,760 tokens.

Llama 4 Scout 17B 16E Instruct is a Mixture-of-Experts model from Meta with 17 billion parameters per expert and 16 experts, activating one expert per token. It supports vision, code generation, tool calling, and 12 languages, making it one of the most versatile models in the Llama 4 family. Scout targets the efficiency-focused segment, offering multimodal capabilities at lower compute cost than dense models of similar quality. Its 10M token context window is among the largest available, and it quantizes well for self-hosted multi-GPU deployments.

At Q2_K quantization (low quality tier), the model weighs 36.85 GB. This fits within the 640 GB of VRAM on NVIDIA H100 640GB, enabling full GPU inference.

Hardware Requirements

Model size	36.85 GB
VRAM available	640 GB
VRAM used	639.4 GB
GPU layers	48 / 48
Context size	3,283,928
Backend	cuda13
Flash attention	Yes

Performance Notes

Mixture-of-Experts

This is a Mixture-of-Experts model with 16 total experts, of which 1 are active per token. Only the active experts need fast GPU access — inactive expert weights can reside in system RAM with minimal impact on generation speed.

Read the MoE deployment guide →

Shorter conversations

Context size has been reduced due to memory constraints. This may affect the model's ability to handle long conversations.

Deploy

Server Local

Prerequisites

Ensure your GPU nodes are prepared with the NVIDIA container toolkit:

ansible-playbook prositronic.infra.nvidia_container_toolkit

Command

helmfile --state-values-file <(curl -s https://www.prositronic.eu/values/llama-4-scout-17b-16e-instruct/q2_k/nvidia-h100-640gb.yaml) apply

Generated values.yaml

/values/llama-4-scout-17b-16e-instruct/q2_k/nvidia-h100-640gb.yaml

Loading values…

Install llama.cpp

Install llama.cpp from the official build instructions for your platform and backend.

Download Model

curl -L -o llama-4-scout-17b-16e-instruct.gguf "https://huggingface.co/unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF/resolve/main/Llama-4-Scout-17B-16E-Instruct-Q2_K.gguf"

Start Server

llama-server \
  -m llama-4-scout-17b-16e-instruct.gguf \
  --n-gpu-layers 48 \
  --ctx-size 3283928 \
  --flash-attn

Verify

curl http://localhost:8080/health

Explore

• Configure a deployment
• Browse models
• Browse hardware

Other quantizations

• Llama 4 Scout 17B 16E Instruct (Q8_0)
• Llama 4 Scout 17B 16E Instruct (Q5_K_M)
• Llama 4 Scout 17B 16E Instruct (Q4_K_M)

Other hardware

• Llama 4 Scout 17B 16E Instruct on CPU Only
• Llama 4 Scout 17B 16E Instruct on NVIDIA A100 80GB
• Llama 4 Scout 17B 16E Instruct on NVIDIA H100

Frequently Asked Questions

How much VRAM does Llama 4 Scout 17B 16E Instruct (Q2_K) need?

The Q2_K quantization of Llama 4 Scout 17B 16E Instruct requires 36.85 GB. All 48 layers fit in the 640 GB of VRAM available on NVIDIA H100 640GB, enabling full GPU acceleration.

Can I run Llama 4 Scout 17B 16E Instruct on NVIDIA H100 640GB?

Yes. NVIDIA H100 640GB provides 640 GB of VRAM, which is enough to run Llama 4 Scout 17B 16E Instruct (Q2_K) with all layers on the GPU for optimal performance.

What is quantization?

Quantization reduces a model's numerical precision from its original floating-point format to a more compact representation. This shrinks the file size and VRAM footprint, making it possible to run large models on consumer hardware. The trade-off is a small reduction in output quality. Q2_K compresses Llama 4 Scout 17B 16E Instruct from its original size down to 36.85 GB.

What quantization should I choose for Llama 4 Scout 17B 16E Instruct?

Q2_K is a low-quality quantization. Higher-quality quants (Q8, Q6) preserve more model accuracy but need more VRAM. Lower quants (Q4, Q3, Q2) reduce VRAM usage at the cost of some quality. Choose based on your available hardware and quality requirements.

What is flash attention and why is it enabled?

Flash attention is a memory-efficient algorithm that speeds up the attention mechanism in transformer models. It reduces VRAM usage during inference by avoiding the materialisation of the full attention matrix. For Llama 4 Scout 17B 16E Instruct on NVIDIA H100 640GB, flash attention is enabled to maximise context length and throughput within the available 640 GB of VRAM.

What is MoE and how does it affect deployment?

Llama 4 Scout 17B 16E Instruct uses a Mixture-of-Experts (MoE) architecture with 16 experts, of which 1 are active per token. This means only a fraction of the model weights are used for each inference step, allowing MoE models to be larger in total parameter count while remaining efficient at inference time.

How do I run Llama 4 Scout 17B 16E Instruct (Q2_K) with Ollama?

Run ollama run llama4:17b-scout-16e-instruct-q2_k to start Llama 4 Scout 17B 16E Instruct (Q2_K). Ollama handles downloading the model weights automatically on first run.

Last updated: March 5, 2026