Llama 4 Maverick 17B 128E Instruct (Q6_K_XL) — 49.3 GBon OVH l40s-1-gpu

Q6_K_XL OVH l40s-1-gpu

Overview

Llama 4 Maverick 17B 128E Instruct is a 396.58B parameter moe language model by Meta, with code, multilingual, tool-calls, vision capabilities. It supports a context window of up to 1,048,576 tokens.

Llama 4 Maverick 17B 128E Instruct is a large-scale Mixture-of-Experts model from Meta with 17 billion parameters per expert and 128 experts, activating one expert per token for a total of approximately 400 billion parameters. It delivers frontier-class performance on vision, code generation, and multilingual tasks across 12 languages. Maverick represents the high-capacity tier of the Llama 4 family, trading higher memory requirements for stronger benchmark results. With a 1M token context window, it requires multi-GPU setups but quantizes down to Q2 levels.

At Q6_K_XL quantization (high quality tier), the model weighs 317.63 GB. This exceeds the 48 GB of VRAM on OVH l40s-1-gpu. Inference is still possible via CPU offload or memory-mapped loading from disk, but expect significantly reduced performance.

The NVIDIA L40S is a datacenter GPU with 48 GB of GDDR6 VRAM and 864 GB/s memory bandwidth. It delivers 362 FP16 TFLOPS with Ada Lovelace architecture. A versatile GPU for AI inference, training, and graphics workloads. Handles quantized models up to 40B parameters comfortably.

Hardware Requirements

Model size	317.63 GB
VRAM available	48 GB
VRAM used	49.3 GB
System RAM	80 GB
Min RAM required	317.6 GB
GPU layers	0 / 48
Context size	262,144
Backend	cuda13
Flash attention	No
Reading from disk	Yes

Performance Notes

Mixture-of-Experts

This is a Mixture-of-Experts model with 128 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 →

Reduced performance

48 of 48 layers will run on CPU, which may reduce inference speed.

Major performance impact

The model size exceeds available VRAM + RAM. Data will be read from disk on demand, which significantly slows inference.

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-maverick-17b-128e-instruct/q6_k_xl/nvidia-l40s.yaml) apply

Generated values.yaml

/values/llama-4-maverick-17b-128e-instruct/q6_k_xl/nvidia-l40s.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-maverick-17b-128e-instruct.gguf "https://huggingface.co/unsloth/Llama-4-Maverick-17B-128E-Instruct-GGUF/resolve/main/UD-Q6_K_XL/Llama-4-Maverick-17B-128E-Instruct-UD-Q6_K_XL-00001-of-00008.gguf"

Start Server

llama-server \
  -m llama-4-maverick-17b-128e-instruct.gguf \
  --n-gpu-layers 0 \
  --ctx-size 262144

Verify

curl http://localhost:8080/health

Explore

• Configure a deployment
• Browse models
• Browse hardware

Other quantizations

• Llama 4 Maverick 17B 128E Instruct (Q8_0)
• Llama 4 Maverick 17B 128E Instruct (Q5_K_M)
• Llama 4 Maverick 17B 128E Instruct (Q4_K_M)

Other hardware

• Llama 4 Maverick 17B 128E Instruct on CPU Only
• Llama 4 Maverick 17B 128E Instruct on NVIDIA A100 80GB
• Llama 4 Maverick 17B 128E Instruct on NVIDIA H100

Frequently Asked Questions

How much VRAM does Llama 4 Maverick 17B 128E Instruct (Q6_K_XL) need?

The Q6_K_XL quantization of Llama 4 Maverick 17B 128E Instruct requires 317.63 GB. The 48 GB of VRAM on OVH l40s-1-gpu is insufficient for GPU layers, so inference runs on CPU.

Can I run Llama 4 Maverick 17B 128E Instruct on OVH l40s-1-gpu?

It is possible but not recommended. OVH l40s-1-gpu does not have enough VRAM to accelerate Llama 4 Maverick 17B 128E Instruct (Q6_K_XL), so inference will rely on CPU and system RAM.

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. Q6_K_XL compresses Llama 4 Maverick 17B 128E Instruct from its original size down to 317.63 GB.

What quantization should I choose for Llama 4 Maverick 17B 128E Instruct?

Q6_K_XL is a high-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.

Why are some layers offloaded to CPU?

OVH l40s-1-gpu has 48 GB of VRAM, but Llama 4 Maverick 17B 128E Instruct (Q6_K_XL) requires approximately 317.63 GB. Only 0 of 48 layers fit in VRAM; the remaining layers run on CPU, which is slower but still functional.

What is MoE and how does it affect deployment?

Llama 4 Maverick 17B 128E Instruct uses a Mixture-of-Experts (MoE) architecture with 128 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 Maverick 17B 128E Instruct (Q6_K_XL) with Ollama?

Run ollama run llama4:17b-maverick-128e-instruct-q6_k_xl to start Llama 4 Maverick 17B 128E Instruct (Q6_K_XL). Ollama handles downloading the model weights automatically on first run.

Last updated: March 5, 2026