Hy4-preview

Hy4-preview

Generates text, code and documents with a Mixture-of-Experts model of 770B parameters that activates 49B and handles up to 1M context. For teams with the hardware to host open weights of this size.

Made by tencent on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Hy4 preview is a flagship Mixture-of-Experts language model from the Tencent Hy Team. The card gives 770B parameters in total, with 49B activated per token, and a context length of 1M. It is aimed at software engineering, office and analysis work, and game development.

Who it suits: It suits teams that run huge open-weight models on their own infrastructure and want one for coding, document work or game prototyping.

Industry | Who uses it | How they use it | Example
Software | Engineers | Plan, debug and verify long development tasks | Work through a front-end feature over many steps
Office work | Analysts and finance teams | Turn messy files into documents, spreadsheets and slides | Build a financial model from scattered files
Games | Game developers | Start from one prompt and refine a prototype over several turns | A playable prototype that they keep improving

Small teams without datacenter-scale hardware can skip hosting it, since the weights alone are huge.

Where to use it:

• Hugging Face Transformers, the library Hugging Face lists for it
• vLLM and SGLang, which the card's contents list as deployment options
• Windows, macOS or Linux, where you can download the files with Python

Use case | Who | What they get back
Software engineering | Developers | Help with planning, debugging and checking long tasks
Office and analysis | Analysts | Documents, spreadsheets and presentations built from many files
Game prototyping | Game developers | A prototype from one prompt that can be refined across turns

How to set it up: You need these first:

• Python
• The huggingface_hub package
• Storage and GPU memory for weights this large, about 1,560 GB in BF16 or about 390 GB at 4-bit as an estimate

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. To serve the model, follow the card's vLLM or SGLang deployment sections.

pip install -U huggingface_hub
hf download tencent/Hy4-preview

How to use it: Download the files, serve them with vLLM or SGLang, then send prompts for coding, analysis or game work. The card text here has no run commands, so follow its Quickstart and Deployment sections for the exact calls.

Example: A developer could give it one prompt describing a small game and ask for a playable prototype. The card says the model works with game engines, so the developer can then keep refining the project over several turns.

For newcomers and experts: Hy4 preview is a huge language model, and its files are too big for an ordinary computer. A realistic first step is to read the card and its Quickstart before downloading anything.

The backbone has 78 layers. The first uses a standard dense FFN. The other 77 are MoE layers with 256 routed experts and 1 shared expert, and each token uses the top 8 routed experts plus the shared one. One native MTP layer supports speculative decoding. Attention is Gated DeepSeek Sparse Attention with IndexCache, and the residual path uses identity Hyper-Connections. The card's 770B leaves out the MTP layer, and Hugging Face lists 780B for the files.

A prompt goes in
A router picks 8 of 256 experts per token
The shared expert adds its output
The model writes the reply

Good to know:

• Size: Hugging Face lists 780B parameters, about 1,560 GB in BF16 and about 390 GB at 4-bit, as an estimate
• Versions: Hugging Face lists 6 quantized and 2 fine-tuned versions made by other people
• Sections: the card's contents also list Finetuning and Quantization, which this text does not include

Title: Parameters in billions
Total backbone: 770 B
Activated per token: 49 B

Title: Weight size in GB
BF16: 1,560 GB
4-bit estimate: 390 GB

• License: apache-2.0
• Cost to run: none named, no API key or account
• Data it reaches: it downloads model files from Hugging Face, and the card text names no other service or account
• Not for: small machines, since the weights alone take about 1,560 GB in BF16

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 780B
• Weights: about 1,560 GB in BF16; about 390 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 11,424
• Likes: 557
• License: apache-2.0
• Quantized versions: 6
• Fine-tunes: 2
• Tags: transformers, safetensors, hy_v4, text-generation, hunyuan, hy4, conversational

Links

• Model: https://huggingface.co/tencent/Hy4-preview
• Files: https://huggingface.co/tencent/Hy4-preview/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download tencent/Hy4-preview

Then follow the usage in its model card. Nothing in this kit runs by itself.