✎ AI News

Tencent Put a 224B Model on Hugging Face, and Only 12B Runs at a Time

Hy4 flash is out as downloadable weights: ~224 billion parameters with ~12 billion active per token, one million tokens of context, text and image input, under the Tencent Hy License.

By AIToolsRecap November 21, 2026 5 min read 127 views
Home › Articles › AI News › Tencent Publishes Hy4 Flash Weights for Local Use
Tencent Put a 224B Model on Hugging Face, and Only 12B Runs at a Time
The short version
● What Tencent published the Hy4 flash weights for local deployment.
● Size About 224 billion parameters, roughly 12 billion active per token.
● Context Up to one million tokens, text and image input.
● Licence Tencent Hy License, on Hugging Face.

Tencent has released the model weights for Hy4 flash, two days after launching it through its own apps and cloud API. It is a sparse mixture-of-experts model of roughly 224 billion parameters with around 12 billion active per token, which is the ratio that makes a model this size realistic to serve.

It takes text and image input and holds up to one million tokens of context. Tencent aims it at ordinary productivity work — research, document preparation, spreadsheets — rather than at benchmark leadership.

What was published

Detail
ModelHy4 flash
Parameters~224 billion, ~12 billion active per token
ContextUp to 1,000,000 tokens
InputText and image
VariantsStandard, FP8, and a base model for further training
LicenceTencent Hy License
Weightshuggingface.co/tencent/Hy4-flash

The base model is the part that matters for anyone doing their own post-training. Releasing a base alongside the instruction-tuned version is a choice plenty of labs skip, and it is the difference between a model you can use and a model you can build on.

The 12-billion active figure is the story

A 224-billion-parameter model sounds out of reach for local deployment. Twelve billion active per token does not. Sparse activation is what makes a model of this size run on hardware a team might actually own, and it is why open-weight releases have moved towards this shape.

Tencent did not publish benchmark figures with the weights, and did not state hardware requirements. Both are worth establishing yourself before committing to it.

How it sits against the other Chinese flagship

Alibaba shipped Qwen4-Max five days ago at the opposite end of the same market: larger, closed, and API-only. We put the two side by side in Hy4 flash versus Qwen4-Max.

FAQ

What is Tencent Hy4 flash?
A text-and-image model of roughly 224 billion parameters with around 12 billion active per token, holding up to one million tokens of context.
Are the Hy4 flash weights open?
They are published on Hugging Face under the Tencent Hy License. Read the licence terms before commercial use.
What variants were released?
The standard model, an FP8 version, and a base model intended for further training.
What is Hy4 flash good at?
Tencent aims it at everyday productivity work: research, document preparation and spreadsheet tasks.
Did Tencent publish benchmarks?
No. No benchmark figures or hardware requirements were published alongside the weights.
Tags
AI NewsGenerative AI2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →