Meta Llama 3.1 8B Instruct (FP32) — 19.4 GBon Hetzner GEX44

FP32 Hetzner GEX44

Overview

Meta Llama 3.1 8B Instruct is a 8B parameter dense language model by Meta, with code, multilingual, tool-calls capabilities. It supports a context window of up to 131,072 tokens.

Meta Llama 3.1 8B Instruct is an 8-billion-parameter dense transformer model from Meta, designed for instruction following, code generation, and multilingual tasks. It offers a strong balance of quality and efficiency in the small-model category, outperforming many 7B-class alternatives on standard benchmarks. The model supports tool calling and eight languages including English, German, and French. With a 128K context window and flash attention support, it runs comfortably on a single consumer GPU at Q4 quantization levels.

At FP32 quantization (full-precision quality tier), the model weighs 29.92 GB. This exceeds the 20 GB of VRAM on Hetzner GEX44. Inference is still possible via CPU offload or memory-mapped loading from disk, but expect significantly reduced performance.

Hardware Requirements

Model size	29.92 GB
VRAM available	20 GB
VRAM used	19.4 GB
System RAM	64 GB
Min RAM required	12.2 GB
GPU layers	19 / 32
Context size	3,197
Backend	cuda13
Flash attention	Yes

Performance Notes

Reduced performance

13 of 32 layers will run on CPU, which may reduce inference speed.

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/meta-llama-3-1-8b-instruct/fp32/nvidia-rtx4000sff.yaml) apply

Generated values.yaml

/values/meta-llama-3-1-8b-instruct/fp32/nvidia-rtx4000sff.yaml

Loading values…

Install llama.cpp

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

Download Model

curl -L -o meta-llama-3-1-8b-instruct.gguf "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/main/Meta-Llama-3.1-8B-Instruct-f32.gguf"

Start Server

llama-server \
  -m meta-llama-3-1-8b-instruct.gguf \
  --n-gpu-layers 19 \
  --ctx-size 3197 \
  --flash-attn

Verify

curl http://localhost:8080/health

Explore

• Configure a deployment
• Browse models
• Browse hardware

Other quantizations

• Meta Llama 3.1 8B Instruct (Q8_0)
• Meta Llama 3.1 8B Instruct (Q5_K_M)
• Meta Llama 3.1 8B Instruct (Q4_K_M)

Other hardware

• Meta Llama 3.1 8B Instruct on CPU Only
• Meta Llama 3.1 8B Instruct on NVIDIA A100 80GB
• Meta Llama 3.1 8B Instruct on NVIDIA H100

Frequently Asked Questions

How much VRAM does Meta Llama 3.1 8B Instruct (FP32) need?

The FP32 quantization of Meta Llama 3.1 8B Instruct requires 29.92 GB. 19 of 32 layers fit in the 20 GB of VRAM on Hetzner GEX44; remaining layers are offloaded to CPU.

Can I run Meta Llama 3.1 8B Instruct on Hetzner GEX44?

Yes, with reduced performance. Hetzner GEX44 can run Meta Llama 3.1 8B Instruct (FP32), but only 19 of 32 layers fit in VRAM. The rest are offloaded to CPU.

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. FP32 compresses Meta Llama 3.1 8B Instruct from its original size down to 29.92 GB.

What quantization should I choose for Meta Llama 3.1 8B Instruct?

FP32 is a full-precision 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 Meta Llama 3.1 8B Instruct on Hetzner GEX44, flash attention is enabled to maximise context length and throughput within the available 20 GB of VRAM.

Why are some layers offloaded to CPU?

Hetzner GEX44 has 20 GB of VRAM, but Meta Llama 3.1 8B Instruct (FP32) requires approximately 29.92 GB. Only 19 of 32 layers fit in VRAM; the remaining layers run on CPU, which is slower but still functional.

How do I run Meta Llama 3.1 8B Instruct (FP32) with Ollama?

Run ollama run llama3.1:8b-instruct-fp32 to start Meta Llama 3.1 8B Instruct (FP32). Ollama handles downloading the model weights automatically on first run.

Last updated: March 5, 2026