DeepSeek Model Download

DeepSeek R1 Series (Reasoning Models, Recommended)

Other Model Series

DeepSeek V3 Quantized Versions (Available on Ollama)

DeepSeek V3 is a 671B-parameter MoE flagship model, and Ollama offers multiple quantized versions. Note: High-precision versions (Q8_0 / FP16) have a default context window of only 4K, requiring manual configuration adjustments.

Important: The Q8_0 and FP16 versions have a default context window of only 4K (instead of 160K), mainly because high-precision models require enormous VRAM at 160K context. For larger context, it is recommended to use the Q4_K_M version, or adjust the num_ctx parameter via a custom Modelfile. The V3 model requires at least ~400GB VRAM/RAM (multi-GPU cluster), and cannot run on a typical personal computer.

What is Q4_K_M quantization?

Q4_K_M is Ollama's default quantization format, belonging to the GGUF quantization series:

• Q4: Indicates 4-bit quantization, compressing the original FP16 (16-bit) weights to 4-bit, reducing model size by about 75%
• K: Indicates the use of K-quant quantization strategy, which uses higher precision (e.g., 6-bit) for important weight layers and lower precision for less important layers
• M: Indicates Medium size, balancing between Q4_K_S (small) and Q4_K_L (large)
• Effect: Significantly reduces VRAM usage and model size with almost no loss in inference quality, making it the officially recommended best choice for Ollama

How to download a specific quantization version?

Ollama supports specifying a specific quantization version by adding the quantization tag after the model name:

# Download Q4_K_M quantized version (default, recommended)
ollama pull deepseek-r1:8b-q4_K_M

# Download smaller Q3_K_M quantized version (slightly lower quality, smaller size)
ollama pull deepseek-r1:8b-q3_K_M

# Download highest precision Q8_0 quantized version (highest quality, largest size)
ollama pull deepseek-r1:8b-q8_0

# If the quantization tag is omitted, Q4_K_M is downloaded by default
ollama pull deepseek-r1:8b

Comparison of Available Quantization Methods

Complete list of GGUF quantization formats from low to high precision (arranged high to low):

Size ratio is based on the FP16 original model. For daily use, Q4_K_M is recommended; if you need the highest quality, choose Q5_K_M or Q6_K.

Method 1: huggingface-cli (Recommended)

Hugging Face's official command-line tool, supports resumable downloads and multi-threaded downloads, best for downloading large models.

# Install huggingface-cli
pip install -U huggingface_hub

# Download full repository
huggingface-cli download deepseek-ai/DeepSeek-R1 --local-dir ./DeepSeek-R1

# Download distilled version (recommended for beginners)
huggingface-cli download deepseek-ai/DeepSeek-R1-Distill-Qwen-7B --local-dir ./R1-Distill-7B

# Exclude certain large files to save space
huggingface-cli download deepseek-ai/DeepSeek-R1 --local-dir ./DeepSeek-R1 --exclude "*.bin"

Method 2: Git Clone + Git LFS

Suitable for developers who need Git version control, can track model updates.

# Install git-lfs
git lfs install

# Clone repository (large files downloaded via LFS)
git clone https://huggingface.co/deepseek-ai/DeepSeek-R1

# If LFS files fail to download, manually pull
cd DeepSeek-R1
git lfs pull

Method 3: Python snapshot_download

Download directly in Python code, suitable for use in Jupyter Notebook or scripts.

from huggingface_hub import snapshot_download

# Download to specified directory
snapshot_download(
repo_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
local_dir="./deepseek-r1-distill",
resume_download=True, # Support resumable downloads
max_workers=4 # Multi-threaded downloads
)

Method 1: Set environment variable (recommended)

After setting, all Hugging Face tools automatically use the mirror:

# Linux / macOS
export HF_ENDPOINT=https://hf-mirror.com

# Windows PowerShell
$env:HF_ENDPOINT = "https://hf-mirror.com"

# Then use huggingface-cli normally to download
huggingface-cli download deepseek-ai/DeepSeek-R1-Distill-Qwen-7B --local-dir ./model

Method 2: Replace URL prefix

Replace huggingface.co with hf-mirror.com:

# Original link
# https://huggingface.co/deepseek-ai/DeepSeek-R1

# Mirror link
# https://hf-mirror.com/deepseek-ai/DeepSeek-R1

# Use git clone with mirror
git clone https://hf-mirror.com/deepseek-ai/DeepSeek-R1

Method 3: Using the mirror in Python

from huggingface_hub import snapshot_download
import os

os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"

snapshot_download(
repo_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
local_dir="./model"
)

Method 1: ModelScope SDK (Recommended)

Use the official ModelScope Python SDK to download, supporting resumable downloads and multi-threaded acceleration. This is the best way to download large models domestically.

# Install ModelScope SDK
pip install modelscope

# Download the full model (automatically from domestic nodes)
from modelscope import snapshot_download

snapshot_download(
model_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
cache_dir="./deepseek-models",
revision="master"
)

# Download V3 full version (requires ~687GB disk space)
snapshot_download(
model_id="deepseek-ai/DeepSeek-V3",
cache_dir="./deepseek-models"
)

Method 2: Git Clone

ModelScope supports the Git protocol, and you can directly use git clone to clone the repository.

# Install git-lfs
git lfs install

# Clone ModelScope repository (extremely fast domestically)
git clone https://www.modelscope.cn/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B.git

# If LFS file download fails, manually pull
cd DeepSeek-R1-Distill-Qwen-7B
git lfs pull

Method 3: Direct Web Download

Visit the ModelScope webpage and click the download button directly on the model page. Suitable for scenarios where you only need to download a single file or a small model.

# Visit ModelScope webpage
# https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B/files
# In the file list, click on a single file to download

GGUF Format Features

• Single-file storage: Model weights, configuration, and tokenizer are all packaged into one .gguf file, making it easy to transfer and manage.
• Built-in quantization: Supports multiple quantization precisions (Q2_K to Q8_0), allowing flexible selection based on hardware resources.
• CPU inference optimization: Deeply optimized for CPU inference, runs without a GPU.
• Memory mapping: Supports mmap() memory mapping for zero-copy loading, extremely fast startup.
• Cross-platform compatibility: Works on Windows, macOS, and Linux without additional dependencies.
• Use cases: Local inference, personal use, low-spec devices, quick deployment.

SafeTensors Format Features

• Safety: Compared to traditional pickle format (.pt/.bin), SafeTensors does not execute arbitrary code, eliminating the risk of malicious code in model files.
• Zero-copy loading: Supports memory mapping, no extra memory copy when loading large models.
• Lazy loading: Supports loading partial weights on demand, no need to load the entire model into memory at once.
• Sharding support: Large models can be split into multiple .safetensors files for distributed storage and loading.
• Framework-agnostic: Readable by major frameworks such as PyTorch, TensorFlow, JAX, etc.
• Use cases: Model training, fine-tuning, research and development, production deployment.