Tencent Opens All Its Hy4 Weights, Adding a License Gate for Large API Providers

The flagship Hy4 followed Hy4 flash onto Hugging Face on November 20, after the cheaper model processed more than 50 trillion tokens on its first day.

WHY IT MATTERS

Cloud platforms and API resellers with over US$50 million in annual revenue now need Tencent's written consent to sell Hy4 access, a condition the Apache-licensed Hy4 preview never carried.

Tencent has published the weights of its flagship Hy4 model, 16 days after launching it as a cloud service and two days after doing the same for the cheaper Hy4 flash. IT Home reported on November 20 that Hy4 is now downloadable on GitHub, Hugging Face, ModelScope and GitCode, together with a system card describing its safety testing. With that release, IT Home says, every model in Tencent’s Hunyuan large language model series is open: Hy3 preview, Hy3, Hy4 preview, Hy4 and Hy4 flash.

Both new releases carry the Tencent Hy License, Version 1.0, a text adapted from the Apache 2.0 open-source license with added commercial terms. A company that sells access to the models as a service needs a separate written license from Tencent if its worldwide revenue exceeds US$50 million in any 12-month period, or if the product using the model has more than 100 million monthly users. The Hy4 preview that Tencent opened in August came under the plain Apache License 2.0. IT Home’s report says the final Hy4 also uses Apache 2.0, but the Hy4 model page and its license file both name the Tencent Hy License.

The tighter terms arrive as usage climbs. Tencent says Hy4 flash processed more than 50 trillion tokens on its first day, the highest first-day usage of any Hunyuan model. It also says the model ranked first on OpenRouter’s daily leaderboard of language models on November 17, according to TechNode. The usage figures come from Tencent and have not been checked by an outside party.

Hy4 flash drew ten times the first-day demand of the Hy4 preview

The Paper, in a report carried on Tencent’s own news app late on November 18, gave the fuller set of Tencent figures. On WorkBuddy, Tencent’s office agent, first-day calls to Hy4 flash were 10 times those to the Hy4 preview, and the model helped complete more than 50 million user tasks. Tencent puts the model’s generation speed at about 200 tokens a second and its peak processing at more than 1 billion tokens a second. Spread evenly over 24 hours, 50 trillion tokens would average about 580 million a second, by EastFrontier’s calculation.

TechNode notes that Hy4 flash has so far avoided the widespread usage limits that followed Hy4’s debut on WorkBuddy on November 4. Prices are unchanged from the November 16 launch, when EastFrontier reported the model’s costs and Tencent had not yet given a date for the weights.

Those weights followed on November 18. Phoenix New Media reported that day that Tencent released three versions: the instruction-tuned Hy4 flash, an FP8 version compressed to need less computing power and memory, and Hy4 flash-base, a pretrained model for further training. TechNode’s report on the weights followed a day later. Tencent’s launch materials gave the model 224 billion parameters, but its Hugging Face model card lists 220 billion, including 64 billion Engram parameters, with 12 billion active for each token.

The new license replaces a regional ban with a size test

Some of Tencent’s earlier open models came with geographic limits. The Tencent Hunyuan Community License attached to Hunyuan-A13B in June 2025 defined its territory as worldwide, excluding the European Union, the United Kingdom and South Korea. The Hy4 flash license contains no territorial carve-out and grants a worldwide, royalty-free copyright license.

The restriction now falls on scale. Section 10 defines a model-as-a-service business as one that gives third parties remote access to a model’s inference or fine-tuning, for example through an API. The revenue test counts a company’s total worldwide gross revenue, whether or not that revenue comes from the model, so almost any large cloud provider would meet it. Purely internal use is exempt, provided neither the model nor its outputs reach third parties.

Companies below both thresholds, and any company running the models only for its own staff, can download, adapt and deploy Hy4 and Hy4 flash without asking Tencent. The largest platforms that would resell them as a service must first agree terms with it.

Outside testing of Hy4’s safety claims can now begin

Hy4 is a mixture-of-experts model with 770 billion parameters, of which 49 billion are active for each token, and a context window of more than 1 million tokens. When it launched, Tencent held back the weights and said it wanted a more responsible release, with attention to cyber and biological risks.

IT Home’s account of the system card says Hy4’s training data included only defensive security material, such as proof-of-concept generation and vulnerability repair, and left out exploitation, reverse engineering and penetration testing. Audits by Tencent’s Xuanwu Lab and Keen Lab found no obvious risk of offensive capability.

Researchers can now test those findings on the weights themselves. The deployment examples on the model page split the full model across eight graphics processors. The first independent benchmark results for Hy4, and any announcement of a cloud platform signing a separate license with Tencent, will show how far the release spreads beyond Tencent’s own services.