SGLang Hugging Face Model Repository IDs | Generated by AI

Home 2026.07

Question: Which Hugging Face repositories can be used with SGLang when it asks for a model name?

Answer:

SGLang loads models directly from the Hugging Face Hub. In most cases, you simply provide the repository ID (owner/model_name), not a URL.

For example:

Qwen/Qwen3-8B
Qwen/Qwen3-32B
meta-llama/Llama-3.1-8B-Instruct
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
google/gemma-3-4b-it
mistralai/Mistral-7B-Instruct-v0.3

or the full URL also works in some UIs:

https://huggingface.co/Qwen/Qwen3-8B

but the repository ID is the standard form.

Models that work well with SGLang

SGLang has excellent support for most modern decoder-only LLMs.

Family Example repo
Qwen Qwen/Qwen3-8B
Llama meta-llama/Llama-3.1-8B-Instruct
Gemma google/gemma-3-4b-it
Mistral mistralai/Mistral-7B-Instruct-v0.3
DeepSeek deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
Phi microsoft/Phi-4-mini-instruct
GLM THUDM/GLM-4-9B-Chat

Models that usually do NOT work

SGLang expects a Hugging Face Transformers model. These generally won’t work directly:

For example:

unsloth/gemma-4-e4b-it-gguf

is a GGUF repository and is intended for llama.cpp, not SGLang.

Instead, use the original Transformers model, e.g.:

google/gemma-3-4b-it

or another supported FP16/BF16/official checkpoint.

Example

If your environment variable is:

MODEL=Qwen/Qwen3-8B

then SGLang downloads the model automatically from Hugging Face:

python -m sglang.launch_server \
    --model-path Qwen/Qwen3-8B

or

MODEL=Qwen/Qwen3-8B
HF_TOKEN=<PLACEHOLDER>

if the model requires authentication.

If you’re using RunPod Serverless

In the “Model” field, simply enter:

Qwen/Qwen3-8B

or

meta-llama/Llama-3.1-8B-Instruct

not a GGUF repository.

References:


Back Donate