DeepSeek V3.1 (Q3_K_M) — 95.4 GBon Hetzner GEX131

Q3_K_M Hetzner GEX131

Overview

DeepSeek V3.1 is a 684.53B parameter moe language model by DeepSeek, with code, multilingual, thinking, tool-calls capabilities. It supports a context window of up to 163,840 tokens.

DeepSeek V3.1 is a 685-billion-parameter Mixture-of-Experts model from DeepSeek, activating 8 of 256 experts per token plus one shared expert. It delivers frontier-level performance on code generation, reasoning, and multilingual tasks while using far fewer active parameters per inference step than comparably sized dense models. The model supports thinking mode, tool calling, and nine languages. With a 160K context window, it requires multi-GPU or distributed setups but quantizes down to Q2 levels for reduced VRAM footprint.

At Q3_K_M quantization (low quality tier), the model weighs 298.46 GB. This exceeds the 96 GB of VRAM on Hetzner GEX131. Inference is still possible via CPU offload or memory-mapped loading from disk, but expect significantly reduced performance.

Hardware Requirements

Model size	298.46 GB
VRAM available	96 GB
VRAM used	95.4 GB
System RAM	256 GB
Min RAM required	293.1 GB
GPU layers	61 / 61
Context size	54,523
Backend	cuda13
Flash attention	Yes

Performance Notes

Mixture-of-Experts

This is a Mixture-of-Experts model with 256 total experts, of which 8 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 →

Minimal performance impact

All expert weights offloaded to CPU — only 8/256 experts are activated per token, so the performance impact is negligible.

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/deepseek-v3-1/q3_k_m/nvidia-rtx6000pro.yaml) apply

Generated values.yaml

/values/deepseek-v3-1/q3_k_m/nvidia-rtx6000pro.yaml

Loading values…

Install llama.cpp

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

Download Model

curl -L -o deepseek-v3-1.gguf "https://huggingface.co/unsloth/DeepSeek-V3.1-GGUF/resolve/main/Q3_K_M/DeepSeek-V3.1-Q3_K_M-00001-of-00007.gguf"

Start Server

llama-server \
  -m deepseek-v3-1.gguf \
  --n-gpu-layers 61 \
  --ctx-size 54523 \
  --flash-attn \
  -ot ".ffn_.*_exps.=CPU"

Verify

curl http://localhost:8080/health

Explore

• Configure a deployment
• Browse models
• Browse hardware

Other quantizations

• DeepSeek V3.1 (Q8_0)
• DeepSeek V3.1 (Q5_K_M)
• DeepSeek V3.1 (Q4_K_M)

Other hardware

• DeepSeek V3.1 on CPU Only
• DeepSeek V3.1 on NVIDIA A100 80GB
• DeepSeek V3.1 on NVIDIA H100

Frequently Asked Questions

How much VRAM does DeepSeek V3.1 (Q3_K_M) need?

The Q3_K_M quantization of DeepSeek V3.1 requires 298.46 GB. All 61 layers fit in the 96 GB of VRAM available on Hetzner GEX131, enabling full GPU acceleration.

Can I run DeepSeek V3.1 on Hetzner GEX131?

Yes. Hetzner GEX131 provides 96 GB of VRAM, which is enough to run DeepSeek V3.1 (Q3_K_M) 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. Q3_K_M compresses DeepSeek V3.1 from its original size down to 298.46 GB.

What quantization should I choose for DeepSeek V3.1?

Q3_K_M 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 DeepSeek V3.1 on Hetzner GEX131, flash attention is enabled to maximise context length and throughput within the available 96 GB of VRAM.

What is MoE and how does it affect deployment?

DeepSeek V3.1 uses a Mixture-of-Experts (MoE) architecture with 256 experts, of which 8 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.

Last updated: March 5, 2026