DeepSeek R1 Distill Qwen 32B (Q8_0) — 15.4 GBon NVIDIA P100

Q8_0 NVIDIA P100

Overview

DeepSeek R1 Distill Qwen 32B is a 32.76B parameter dense language model by DeepSeek, with code, multilingual, thinking, tool-calls capabilities. It supports a context window of up to 131,072 tokens.

DeepSeek R1 Distill Qwen 32B is a 32.76-billion-parameter dense transformer from DeepSeek, distilled from the larger R1 reasoning model into a Qwen-based architecture. It excels at chain-of-thought reasoning, code generation, and multilingual tasks with built-in thinking capabilities. Compared to standard 30B-class instruct models, it provides stronger logical and mathematical reasoning. The model supports nine languages and a 128K context window, making it suitable for developers and researchers who need reasoning-focused inference on mid-range GPU setups.

At Q8_0 quantization (high quality tier), the model weighs 32.43 GB. This exceeds the 16 GB of VRAM on NVIDIA P100. Inference is still possible via CPU offload or memory-mapped loading from disk, but expect significantly reduced performance.

Hardware Requirements

Model size	32.43 GB
VRAM available	16 GB
VRAM used	15.4 GB
Min RAM required	18.7 GB
GPU layers	27 / 64
Context size	3,105
Backend	cuda12
Flash attention	Yes

Performance Notes

Reduced performance

37 of 64 layers run on CPU. Your system needs at least 18.7 GB RAM. If system RAM is insufficient, data will be read from disk on demand, with significant performance impact.

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-r1-distill-qwen-32b/q8_0/nvidia-p100.yaml) apply

Generated values.yaml

/values/deepseek-r1-distill-qwen-32b/q8_0/nvidia-p100.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-r1-distill-qwen-32b.gguf "https://huggingface.co/unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF/resolve/main/DeepSeek-R1-Distill-Qwen-32B-Q8_0.gguf"

Start Server

llama-server \
  -m deepseek-r1-distill-qwen-32b.gguf \
  --n-gpu-layers 27 \
  --ctx-size 3105 \
  --flash-attn

Verify

curl http://localhost:8080/health

Explore

• Configure a deployment
• Browse models
• Browse hardware

Other quantizations

• DeepSeek R1 Distill Qwen 32B (FP16)
• DeepSeek R1 Distill Qwen 32B (Q5_K_M)
• DeepSeek R1 Distill Qwen 32B (Q4_K_M)

Other hardware

• DeepSeek R1 Distill Qwen 32B on CPU Only
• DeepSeek R1 Distill Qwen 32B on NVIDIA A100 80GB
• DeepSeek R1 Distill Qwen 32B on NVIDIA H100

Frequently Asked Questions

How much VRAM does DeepSeek R1 Distill Qwen 32B (Q8_0) need?

The Q8_0 quantization of DeepSeek R1 Distill Qwen 32B requires 32.43 GB. 27 of 64 layers fit in the 16 GB of VRAM on NVIDIA P100; remaining layers are offloaded to CPU.

Can I run DeepSeek R1 Distill Qwen 32B on NVIDIA P100?

Yes, with reduced performance. NVIDIA P100 can run DeepSeek R1 Distill Qwen 32B (Q8_0), but only 27 of 64 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. Q8_0 compresses DeepSeek R1 Distill Qwen 32B from its original size down to 32.43 GB.

What quantization should I choose for DeepSeek R1 Distill Qwen 32B?

Q8_0 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.

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 R1 Distill Qwen 32B on NVIDIA P100, flash attention is enabled to maximise context length and throughput within the available 16 GB of VRAM.

Why are some layers offloaded to CPU?

NVIDIA P100 has 16 GB of VRAM, but DeepSeek R1 Distill Qwen 32B (Q8_0) requires approximately 32.43 GB. Only 27 of 64 layers fit in VRAM; the remaining layers run on CPU, which is slower but still functional.

How do I run DeepSeek R1 Distill Qwen 32B (Q8_0) with Ollama?

Run ollama run deepseek-r1:32b-qwen-distill-q8_0 to start DeepSeek R1 Distill Qwen 32B (Q8_0). Ollama handles downloading the model weights automatically on first run.

Last updated: March 5, 2026