> ## Documentation Index
> Fetch the complete documentation index at: https://docs.sglang.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Hy4-flash

> Deploy Tencent Hy4-flash with SGLang — launch recipes for the ~220B Mixture-of-Experts model with sparse + sliding-window attention and Engram memory, in BF16 and FP8 on H200, B300 and GB300.

## Deployment

> [accordion] Install SGLang
>
> For all methods and hardware platforms, see the [official SGLang installation guide](https://docs.sglang.io/docs/get-started/install.md). The two paths below match the **Python / Docker** toggle in the command panel.
>
> > [tab] Docker
> >
> > ```bash Command theme={null}
> > docker pull lmsysorg/sglang:hy4-flash
> > ```
> >
> > For how to launch the image, see [Install → Method 3: Using Docker](https://docs.sglang.io/docs/get-started/install.md#method-3-using-docker), substituting the inner `sglang serve ...` with what the command generator below produces.
>
> > [tab] Python (pip / uv)
> >
> > Hy4-flash support is not in a released SGLang wheel yet. Install SGLang from source on the model-support branch ([sgl-project/sglang#51159](https://github.com/sgl-project/sglang/pull/51159)):
> >
> > ```bash Command theme={null}
> > git clone -b hy4_flash_squash https://github.com/sgl-project/sglang.git
> > cd sglang
> > pip install --upgrade pip
> > pip install -e "python"
> > ```
> >
> > Then run the **Python** output of the command panel below in that environment.

Pick your hardware + recipe to generate the launch command.

## Playground

The Playground is where you experiment with **SGLang features beyond the verified matrix**. The Deploy panel above only emits combinations the SGLang team has signed off on; the Playground lets you turn on prefill/decode disaggregation and hierarchical KV caching (HiCache) on top of whichever cell the Deploy panel is currently showing.

## 1. Model Introduction

**Hy4-flash** is Tencent's \~220B-parameter (\~12B active) Mixture-of-Experts model with 512 routed experts, interleaving DeepSeek-style sparse MLA attention with sliding-window attention and adding Engram n-gram memory tables.

**Available models:**

* [tencent/Hy4-flash](https://huggingface.co/tencent/Hy4-flash) — BF16
* [tencent/Hy4-flash-FP8](https://huggingface.co/tencent/Hy4-flash-FP8) — FP8 (e4m3, 128×128 block scales)

## 2. Advanced Usage

### 2.1 HiCache (Hierarchical KV Caching)

Hy4-flash supports SGLang's hierarchical KV cache, including the host-memory (L2) tier and storage (L3) backends. Toggle it in the **HiCache** card of the [Playground above](#playground) to add the flags to your launch command.

### 2.2 PD Disaggregation

Prefill and decode can run on separate SGLang instances. Pick a role and transfer backend in the **PD Disagg** card of the [Playground above](#playground); it also generates the `sgl-router` command that fronts both roles.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.