⚡ Global Anycast Edge Routing · Full HTTP Range Resumable · Zero Token Storage

Hugging Face Mirror Gateway

High-performance, stable model weights & datasets downloading for AI developers, researchers, and universities worldwide. Direct streaming reverse proxy for backend download interfaces. Transparent token pass-through with zero credentials stored.

📚 Explore AI Engineering Articles (32 Guides) ⚡ 5-Minute Quick Start Configuration

100%

Official SHA256 Verification

HTTP Range

Full Resumable Downloads

0 Logging

Zero Token Retention

10Gbps+

Backbone Streaming Proxy

🚀 Quick Start: One Environment Variable to Accelerate

Compatible with official huggingface-cli, Transformers Python SDK, Git LFS, and Docker containers.

Set HF_ENDPOINT in your terminal, then invoke huggingface-cli:

Linux / macOS (Bash / Zsh):

export HF_ENDPOINT="https://hf-mirror.net"
huggingface-cli download --resume-download deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --local-dir DeepSeek-R1-32B

Windows PowerShell:

$env:HF_ENDPOINT = "https://hf-mirror.net"
huggingface-cli download --resume-download deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --local-dir DeepSeek-R1-32B

Set the environment variable in your Python script before importing Hugging Face libraries:

Python Script (Transformers / snapshot_download):

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

from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Qwen/Qwen2.5-7B-Instruct"
snapshot_download(repo_id=model_id, local_dir="./Qwen2.5-7B")

tokenizer = AutoTokenizer.from_pretrained("./Qwen2.5-7B")
model = AutoModelForCausalLM.from_pretrained("./Qwen2.5-7B")

Directly replace huggingface.co with hf-mirror.net:

Git Clone & LFS Commands:

git lfs install
# Optional: skip large blobs during clone, then pull with mirror
GIT_LFS_SKIP_SMUDGE=1 git clone https://hf-mirror.net/openai/whisper-large-v3.git
cd whisper-large-v3
git lfs pull

Inject HF_ENDPOINT into Docker or docker-compose:

Docker Run Command:

docker run -d --gpus all -p 8000:8000 \
  -e HF_ENDPOINT="https://hf-mirror.net" \
  -v /data/hf_cache:/root/.cache/huggingface \
  vllm/vllm-openai:latest \
  --model Qwen/Qwen2.5-7B-Instruct

🔒 Gated Repo? Set export HF_TOKEN="hf_xxx" to authenticate.

View Detailed Quick Start Manual →

🔥 Trending Models

Popular Open-Source Models

One-click commands to pull current top-ranking LLMs, Multimodal, and Vision weights

Read All 32 Engineering Guides →

Reasoning LLM deepseek-ai

DeepSeek-R1-Distill-Qwen-32B

Leading open-source reasoning model driven by pure reinforcement learning (RL) with chain-of-thought logic.

huggingface-cli download deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --local-dir R1-32B

Flagship LLM Qwen

Qwen2.5-72B-Instruct

Flagship model with 128K context window, superior coding, math reasoning, and multilingual comprehension.

huggingface-cli download Qwen/Qwen2.5-72B-Instruct --local-dir Qwen-72B

Gated Model meta-llama

Llama-3.3-70B-Instruct

Meta's state-of-the-art open-weights model matching 405B capabilities with streamlined 70B compute footprint.

huggingface-cli download meta-llama/Llama-3.3-70B-Instruct --local-dir Llama-3.3-70B

Diffusion DiT black-forest-labs

FLUX.1-dev

12B DiT text-to-image foundation model delivering stunning visual quality, anatomy accuracy, and typography rendering.

huggingface-cli download black-forest-labs/FLUX.1-dev --local-dir FLUX.1-dev

Coding Assistant Qwen

Qwen2.5-Coder-32B-Instruct

Top-tier open-source coding model with 128K context, repository comprehension, and native Continue plugin compatibility.

huggingface-cli download Qwen/Qwen2.5-Coder-32B-Instruct --local-dir Coder-32B

Speech ASR OpenAI

whisper-large-v3-turbo

Lightweight speech recognition model offering 8x faster inference speed and low VRAM footprint with high multilingual precision.

huggingface-cli download openai/whisper-large-v3-turbo --local-dir whisper-turbo

🌐 How It Works

HF-Mirror Gateway Architecture & Data Topology

Traditional international connections frequently drop large files due to cross-border packet loss and TCP window collapse. HF-Mirror utilizes an intelligent, layered proxy pipeline:

1. Client Request

The local CLI or SDK sends resolve requests to hf-mirror.net. The gateway resolves repo metadata and routes requests to high-bandwidth edge backbones.

2. Zero-Copy Streaming

Data streams are forwarded on the fly via high-speed transit with zero disk caching, ensuring maximum download speeds up to your local physical bandwidth limit.

3. Token Pass-Through

Authorization headers (Bearer Token) are forwarded directly to official verification endpoints without intermediate logging or storage.

⚖️ Performance Benchmark

Direct Connection vs HF-Mirror Acceleration

Comparing download reliability and throughput when pulling modern 30GB+ large language models:

Dimension	Direct Connection	HF-Mirror Gateway
Peak Speed	100KB/s ~ 2MB/s (Fluctuating)	20MB/s ~ 100MB/s+ (Max Out Wire)
Stability	Frequent Connection Reset (EOF)	Dedicated High-Availability Backbone
Resume Support	Broken pipe restarts from 0%	Native HTTP Range partial chunks
Gated Repo Tokens	Manual VPN / Proxy config needed	Transparent pass-through, zero config