VRAMGlass contains affiliate links. As an eBay Partner, we may be compensated if you make a purchase.As an eBay Partner, we may be compensated.Learn more

GPUs tracked57Price history49 daysPrice points (24h)43,400 pointsBiggest 24h dropGeForce RTX 3060 12GB-11.1%Biggest 24h riseCMP 170HX 40GB+19.8%Updated
$ USD

How to Deploy DeepSeek R1 Qwen3 8B Locally

DeepSeek R1 Qwen3 8B is Qwen3 8B trained on R1-0528's reasoning traces, and it is the model Ollama pulls when you simply run deepseek-r1. At 4-bit with 8K of context it takes about 7.3 GB, so an 8GB card can run it. It is a clear step up from the original Qwen3 8B at math, but the trade-off is that it writes out a stretch of thinking every time. For a new setup, we now recommend Qwen3.5 9B, whose thinking can be switched on and off.

DeepSeek R1 Qwen3 8B details

  • Min VRAM at 4-bit8 GB · 8K
  • Parameters8.2B
  • Max context128K
Architecture
Dense
License
MIT
Launch date
May 29, 2025
Vendor
DeepSeek
Hugging Face repo
deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
GGUF repo
unsloth/DeepSeek-R1-0528-Qwen3-8B-GGUF
MLX repo (Apple)
mlx-community/DeepSeek-R1-0528-Qwen3-8B-4bit
Ollama tag
deepseek-r1:8b
Revision
0528
Layers
36
Hidden size
4,096
KV heads
8
Head dim
128

How much VRAM does DeepSeek R1 Qwen3 8B need for local deployment?

QuantizationWeightsGB8K contextGB32K contextGB128K context (max)GB
1-bit2.4GGUF4.78.322.8
2-bit3.5GGUF5.89.423.9
3-bit4.3GGUF6.610.224.7
4-bit5GGUF7.310.925.4
5-bit5.9GGUF8.211.826.3
8-bit8.7GGUF1114.629.1
BF16/FP1616.4GGUF18.722.336.8

Full calculation rules are on the Methodology page.

Can a single GPU run DeepSeek R1 Qwen3 8B?

Yes, on almost any discrete GPU; the 4-bit build fits even an 8GB card. What you really need to plan around is context. It writes a long chain of thought before every answer, and on hard problems that reasoning can run to tens of thousands of tokens, so leave at least 32K or it gets cut off mid-thought. By that standard, a 16GB card like the RTX 5060 Ti 16GB can go straight to 8-bit and still fit 32K, with quality close to the original, while a 24GB card runs 8-bit out to about 105K, at roughly 70 tokens/s on an RTX 3090. On 12GB cards, use 4-bit. 2-bit saves too little space to be worth it on a model this small.

eBay
DevicePriceSpeed t/sMax context
Tesla V100 16GB16 GB$37024 listings105.273K
GeForce RTX 5060 Ti 16GB16 GB$86958 listings56.673K
Tesla V100 32GB32 GB$89518 listings105.2128K
Radeon RX 7900 XTX24 GB$1,25029 listings111.1128K
GeForce RTX 309024 GB$1,799107 listings108.8128K
GeForce RTX 409024 GB$3,75868 listings115.8128K
Ryzen AI Max+ 395 128GBWhole system · 128 GB$4,46420 listings33.5128K
RTX A600048 GB$4,89916 listings91.8128K
Mac Studio M5 Ultra 96GBWhole system · 96 GB$7,27412 listings136.3128K
MacBook Pro M5 Max 128GBWhole system · 128 GB$7,54920 listings75.3128K
DGX Spark 128GBWhole system · 128 GB$8,24022 listings35.6128K
GeForce RTX 509032 GB$8,395123 listings182.8128K
RTX PRO 600096 GB$22,50011 listings182.8128K

Only common GPUs and systems are listed, one set for models under 50B and another for larger ones, sorted by price from low to high.

What to know before running DeepSeek R1 Qwen3 8B locally

  1. It isn't the full R1

    Despite R1 in the name, it is not the same thing as the full 671B R1. Its architecture is identical to Qwen3 8B; DeepSeek simply post-trained Qwen3 8B on R1-0528's reasoning traces. DeepSeek specifically points out that its tokenizer configuration follows R1-0528, so when you deploy or convert it yourself, take the config files from DeepSeek's repository rather than substituting the original Qwen3 ones.

  2. Which quantization to pick

    If you have the VRAM, use 8-bit: the weights are 8.7 GB, and with 32K of context the total is about 14.6 GB, which fits on a 16GB card. On a 12GB card, use 4-bit (about 10.9 GB at 32K). 2-bit is only 1.5 GB smaller than 4-bit, and for an 8B model that saving isn't worth it.

  3. Official recommended settings

    DeepSeek's website and app use temperature 0.6, and its evaluations use top_p 0.95. Unlike the 14B and 32B distills from January 2025, the 0528 version supports system prompts and no longer needs the thinking tag added manually at the start of the output.

  4. Don't run hard problems at the default context

    Default context is often too short, so long thinking gets cut off. Ollama sets its default based on VRAM: only 4K below 24GB, and 32K from 24 to 48GB. The llama-server command on this page uses 8K (-c 8192). For math and coding problems, raise -c to 32768 or more in llama-server; in Ollama, increase it with the OLLAMA_CONTEXT_LENGTH environment variable or /set parameter num_ctx.

Launch commands

Ollama

  • Windows
  • macOS
  • Linux
Terminal
ollama run deepseek-r1:8b

llama.cpp

  • NVIDIA
  • AMD
  • Mac
  • CPU
Terminal
llama-server -hf unsloth/DeepSeek-R1-0528-Qwen3-8B-GGUF:Q4_K_M -c 8192

vLLM

  • Linux
  • NVIDIA
  • AMD
Terminal
vllm serve deepseek-ai/DeepSeek-R1-0528-Qwen3-8B

MLX

  • Apple Silicon

On a Mac, pip install mlx-lm first; this command downloads the MLX weights from Hugging Face and starts an OpenAI-compatible server.

Terminal
mlx_lm.server --model mlx-community/DeepSeek-R1-0528-Qwen3-8B-4bit

File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology

Other models in the DeepSeek family

ModelTotal paramsActive params4-bit size
DeepSeek R1 Distill Qwen 14BDense14.8B14.8B9 GB
DeepSeek R1 Distill Qwen 32BDense32.8B32.8B19.9 GB
DeepSeek V4 FlashMoE284B13B155.1 GB
DeepSeek V3.2MoE671B37B405.4 GB
DeepSeek V4.1 FlashMoE763B8 / 16B444.7 GB
DeepSeek V4 ProMoE1,600B49B849.7 GB

About DeepSeek