Where to Find Distilled Versions of DeepSeek-R1 Models: Complete Download Guide

All six distilled versions of DeepSeek-R1 are openly hosted on Hugging Face and documented in the repository's README.md under the "DeepSeek-R1-Distill Models" section.

The distilled versions of DeepSeek-R1 models provide compact, efficient alternatives to the full 671B parameter model while retaining strong reasoning capabilities. According to the deepseek-ai/DeepSeek-R1 source code, these distilled checkpoints are openly distributed through Hugging Face with complete documentation for deployment and inference.

What Are Distilled DeepSeek-R1 Models?

Distilled models are compact versions created by transferring knowledge from the full DeepSeek-R1 model to smaller base architectures. The repository provides six distilled variants built on Qwen2.5-Math and Llama-3.1/3.3 foundations, ranging from 1.5B to 70B parameters.

Complete List of Distilled Models

The repository's README.md (lines 81-92) contains a comprehensive table listing all available distilled checkpoints. Each model maintains compatibility with its upstream base architecture while incorporating DeepSeek-R1's reasoning patterns.

Distilled Model	Base Model	Hugging Face Download
DeepSeek-R1-Distill-Qwen-1.5B	Qwen2.5-Math-1.5B	Download
DeepSeek-R1-Distill-Qwen-7B	Qwen2.5-Math-7B	Download
DeepSeek-R1-Distill-Llama-8B	Llama-3.1-8B	Download
DeepSeek-R1-Distill-Qwen-14B	Qwen2.5-14B	Download
DeepSeek-R1-Distill-Qwen-32B	Qwen2.5-32B	Download
DeepSeek-R1-Distill-Llama-70B	Llama-3.3-70B-Instruct	Download

How to Download and Run Distilled Models

All distilled versions of DeepSeek-R1 models can be deployed using standard inference engines. The repository provides specific launch commands for vLLM and SGLang in the README.md documentation.

Download from Hugging Face

Access the models directly using the Hugging Face Hub:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

Deploy with vLLM

The README.md (lines 174-177) provides the following vLLM command for serving the 32B parameter model:

vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B \
      --tensor-parallel-size 2 \
      --max-model-len 32768 \
      --enforce-eager

Deploy with SGLang

For SGLang deployment, use the configuration shown in README.md lines 180-184:

python3 -m sglang.launch_server \
       --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B \
       --trust-remote-code \
       --tp 2

API Integration

When using the OpenAI-compatible API, the README recommends setting the temperature between 0.5 and 0.7 for optimal reasoning performance:

import openai

client = openai.OpenAI(
    base_url="https://api.deepseek.com/v1",
    api_key="YOUR_API_KEY"
)

response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
    messages=[{"role": "user", "content": "Explain the prime number theorem."}],
    temperature=0.6,
    max_tokens=1024,
)

Key Files in the Repository

Understanding the repository structure helps locate additional technical details about the distilled versions of DeepSeek-R1 models.

File	Purpose
README.md	Primary documentation containing the distilled model table (lines 81-92) and deployment examples (lines 174-184)
DeepSeek_R1.pdf	Research paper detailing the distillation methodology and training pipeline
.github/workflows/*	CI/CD infrastructure for repository maintenance

Summary

• All six distilled versions of DeepSeek-R1 models are openly available on Hugging Face under the deepseek-ai organization.
• Model sizes range from 1.5B to 70B parameters, built on Qwen2.5-Math and Llama-3.1/3.3 architectures.
• Deployment options include direct Hugging Face Transformers, vLLM (lines 174-177), and SGLang (lines 180-184).
• Optimal inference parameters include temperatures between 0.5 and 0.7 for reasoning tasks.

Frequently Asked Questions

What is the difference between DeepSeek-R1 and its distilled versions?

The full DeepSeek-R1 is a 671B parameter mixture-of-experts (MoE) model requiring substantial computational resources. The distilled versions of DeepSeek-R1 models transfer the reasoning capabilities to smaller architectures (1.5B to 70B parameters) based on Qwen and Llama, making them accessible for local deployment while maintaining strong reasoning performance.

Which distilled model should I use for production?

For production environments, the DeepSeek-R1-Distill-Qwen-32B or DeepSeek-R1-Distill-Llama-70B offer the best balance of performance and resource requirements. The 32B Qwen variant can be served efficiently using vLLM with tensor parallelism (as shown in README.md lines 174-177), while the 70B Llama variant provides the highest accuracy for complex reasoning tasks.

Do I need a GPU to run distilled DeepSeek-R1 models?

Yes, GPU acceleration is required for practical inference, though the specific requirements vary by model size. The smaller variants (1.5B and 7B) can run on consumer GPUs with 8-16GB VRAM, while larger models like the 32B and 70B variants require multiple high-memory GPUs (e.g., A100 or H100) configured with tensor parallelism as demonstrated in the vLLM and SGLang deployment examples.

Where can I find the technical details about the distillation process?

The complete technical methodology is documented in the DeepSeek_R1.pdf file within the repository. This research paper describes how knowledge was transferred from the full 671B parameter model to the smaller Qwen and Llama architectures. Additionally, the README.md file (lines 81-92) provides the practical specifications and Hugging Face links for each distilled checkpoint.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}