In short: Tencent has published the weights of its 770-billion-parameter Hy4 model. Its new licence applies worldwide, unlike some earlier Tencent models that excluded the UK and EU, but large API resellers need Tencent's permission. Published 21 Nov 2026 by the AI Decoded desk; news dated 20 Nov 2026.

Tencent has published the weights of Hy4, its flagship large language model, 16 days after launching it as a cloud service. The release on 20 November means every model in Tencent's Hy series, previously known as Hunyuan, can now be downloaded. For British developers the small print matters as much as the model: the new licence covers the UK, which some of Tencent's earlier open models did not.

What Tencent released

Hy4 first appeared on 4 November on Tencent's own services. The weights, the trained files that let anyone run and modify a model, are now on Hugging Face, ModelScope and GitCode in three versions: the instruction-tuned Hy4, a compressed FP8 version that needs less memory, and a pre-trained base model for further training. Tencent has also published a system card describing its safety testing.

According to the model card, Hy4 is a mixture-of-experts model with 770 billion parameters, of which 49 billion are active for each token, plus a small vision encoder so it can read images and video. It handles up to about one million tokens of context. Tencent credits DeepSeek and Z.ai's GLM models as inspiration for its attention design, a sign of how openly Chinese labs now build on each other's published work.

A cheaper sibling, Hy4 flash, launched on 16 November and had its weights released on 18 November. EastFrontier reported that its Hugging Face card lists 220 billion parameters with 12 billion active per token. Tencent says Hy4 flash processed more than 50 trillion tokens on its first day, a figure that has not been independently checked.

What Tencent claims

The model card describes work on computer use, game development, finance and office tasks. In an evaluation with 213 finance professionals across 579 real-world tasks, Tencent says Hy4 averaged 3.04 out of 4, ahead of DeepSeek-V4.1-Flash-1017 on 2.97 and Kimi K3.1 Pro and GLM-5.3 on 2.68 each. It also reports cutting its hallucination rate, how often the model states things that are false, from 15.76% in the earlier Hy4 preview to 5.45% in internal tests.

Tencent is unusually frank about the limits. The card says that "a notable gap remains between it and top proprietary models", particularly on research-level maths, long coding tasks and advanced visual questions.

On safety, EastFrontier, citing IT Home's account of the system card, reported that Hy4's training data included only defensive security material, such as vulnerability repair, and left out exploitation and penetration testing, and that audits by Tencent's own security labs found no obvious offensive capability. With the weights public, outside researchers can now test that claim directly.

The licence: worldwide, with a size test

Hy4 and Hy4 flash use the new Tencent Hy License, Version 1.0. It is adapted from the widely used Apache 2.0 open-source licence and grants a worldwide, royalty-free right to use, modify and redistribute the models.

The catch is in section 10. A company that offers the models to others as a service, for example through an API, needs a separate written licence from Tencent if its worldwide revenue exceeds $50 million in any 12-month period, or if the product using the model has more than 100 million monthly users. That revenue test counts all of a company's revenue, not just income from the model, so almost any large cloud provider would need to agree terms. Purely internal use is exempt, as long as neither the model nor its outputs reach third parties. The Hy4 preview released in August came under the plain Apache 2.0 licence, without these conditions, EastFrontier noted.

Why it matters for the UK

Some of Tencent's earlier open models came with a community licence that expressly did not apply in the European Union, the United Kingdom or South Korea. As EastFrontier pointed out, the Hunyuan-A13B model released in June 2025 defined its permitted territory as worldwide excluding those places. The Hy4 flash licence contains no such carve-out.

In practice, that means UK start-ups, universities and companies below the size thresholds can download, adapt and deploy Hy4 without asking Tencent, and larger UK firms can use it internally. Big platforms that want to resell access still need Tencent's agreement. As with any open model, organisations should read the licence in full and take their own advice before building a product on it.

What to watch

The next tests are independent benchmark results for Hy4 and whether any major cloud platform signs a separate licence with Tencent. Running the full model is not trivial either: Tencent's own deployment examples spread it across eight graphics processors.

FAQs

Can I download Tencent's Hy4 model?

Yes. Since 20 November 2026 the Hy4 weights, an FP8 version and a base model have been available on Hugging Face, ModelScope and GitCode under the Tencent Hy License, Version 1.0.

Can UK companies use Hy4?

The Hy4 licence grants rights worldwide and, unlike some earlier Tencent licences, does not exclude the UK. Companies offering the model as a service need Tencent's written permission if their worldwide revenue exceeds $50 million a year or the product has more than 100 million monthly users. Purely internal use is exempt.

How does Hy4 compare with the best AI models?

Tencent says Hy4 leads some open-weight rivals in its own finance evaluation, but its model card acknowledges a notable gap to the top closed models on hard maths, long coding tasks and advanced visual questions. Independent results are not yet available.

Sources

Follow AI Decoded on Substack

Weekly plain-English AI news and verified UK events.