The Daily Inference
AI & Technology · News

Free to Download, Not Free to Serve at Scale: Tencent's Hy4 Licence

Developers can download, adapt and redistribute Tencent's model. But large businesses offering general-purpose model access may need separate written permission, and modifying the weights does not remove that condition.

Follows Tencent Cut Hacking From Its AI Training. Here Is What Stayed..

Tencent has released downloadable weights for its Hy4 model under a conditional licence that lets developers build on it while reserving approval rights over certain large service businesses. [1][2][3][4]

The release includes instruction-tuned weights, an FP8 version that stores weights at lower numerical precision, and a pretrained base checkpoint. Tencent's official model page also includes a System Card section. All three repositories identify the Tencent Hy License as their governing licence. [1][2][3]

The licence shows which rights Tencent keeps after handing them over. [4]

Tencent Hy License 1.0 says it derives from Apache 2.0 but is not Apache 2.0, and has no endorsement from or affiliation with the Apache Software Foundation. It grants royalty-free rights to reproduce, modify, distribute and sublicense the model, subject to its conditions. [4]

The commercial gate applies to what the licence calls "model-as-a-service": giving third parties remote access to model inference or fine-tuning, including through an API. That definition, rather than commercial use alone, determines who faces the additional permission requirement. [4]

Two thresholds, one defined business

A model-as-a-service operator needs a separate written licence from Tencent if its aggregate worldwide gross revenue exceeds $50 million in any consecutive 12-month period. The calculation includes revenue unrelated to Hy4. [4]

That is a company-wide revenue test, not a bill for money earned from this particular model. A qualifying operator cannot avoid the condition simply because its Hy4 service is small or has yet to earn much. [4]

The second threshold concerns audience size. An operator must obtain separate permission before using the model for a product or service with more than 100 million monthly active users in any calendar month. The revenue and user tests are separate routes to the same requirement. [4]

But neither number turns every large commercial user into a licensing customer. The operator must fall within the licence's definition of model-as-a-service. Tencent expressly excludes end-user products confined to specific features or workflows that do not provide general-purpose model access. [4]

The licence also excludes businesses that merely relay requests to an independently operated third-party model service. Passing a request along is treated differently from operating the underlying model service. [4]

For builders, the first question is therefore what their product lets customers do. A bounded feature and a general-purpose model endpoint can face different obligations under the same licence. Product design becomes part of the legal decision, alongside revenue and audience size.

There is an internal-use exception, but its wording is tight. The Section 10.2 requirements do not apply to solely internal use only when the model, its derivatives and their outputs are all kept from third parties. Keeping the weights inside the company is insufficient if the outputs go outside it. [4]

Tencent has opened a broad path for development while keeping a permission gate where a large operator offers the model's capabilities to others.

Modified weights carry the terms

Downloading the files gives developers the practical ability to run and change Hy4. Tencent documents deployment through software including vLLM and SGLang, tools that let people run models on their own hardware, and provides the pretrained base as a distinct starting point from the instruction-tuned model. [1][3]

The licence follows those changes. It defines fine-tuned, quantized, distilled, merged and pruned weights as "derivative works", according to the published text. Those are different ways to adapt or reduce a model, and none removes the commercial conditions. [4]

Redistributors must supply a copy of the licence, mark modified files and preserve applicable notices. They may apply their own terms to their work, but those terms cannot remove or conflict with Section 10's additional commercial provisions. [4]

That makes the release useful without making it unrestricted. A developer can improve the model and distribute the result, while the next operator still has to check whether its service crosses Tencent's permission thresholds. [4]

Other rights remain separate too. The licence grants no general right to Tencent's trade names, trademarks, service marks or product names, beyond reasonable use to describe the model's origin and reproduce notices. It supplies the work "as-is", disclaims warranties and limits contributors' liability, subject to applicable law or a written agreement. [4]

The download is not a support contract, and permission to use the model is not permission to market a product as Tencent's own.

A large model, with checks still ahead

Tencent describes Hy4 as a multimodal model capable of understanding text, images and video. Its language backbone has 770 billion parameters, with 49 billion active for each text token, according to the model card. Parameters are the numerical values that shape a model's responses. [1]

The listed context length is 1,048,576 tokens, the units of text a model processes. Tencent also says the base model encountered substantial synthetic instruction data, machine-generated example tasks, during an intermediate training stage it calls mid-training. It warns that this may distort some few-shot benchmarks, which test performance after giving a model a small number of examples. [1]

The System Card's detailed safety tests and conclusions have not been independently verified here. The release does not establish publication of the full training data or complete training process, and the licence describes no technical mechanism for Tencent to switch off downloaded copies remotely. [1][4]

The next practical test is deployment: whether independent users can run these checkpoints economically, reproduce Tencent's assessments and establish which products require separate permission. For service builders, the immediate dividing line is already written down: general-purpose access, worldwide revenue and monthly users. [4]

The story so far 2 pieces since November 15, 2026, in order

Topics: Tencent · Hugging Face · AI business · AI safety

Every edition in brief, three times a day, on our Telegram channel, on Bluesky and on Threads.

Sources
  1. tencent/Hy4 · Hugging Face huggingface.co
  2. tencent/Hy4-FP8 · Hugging Face huggingface.co
  3. tencent/Hy4-base · Hugging Face huggingface.co
  4. huggingface.co huggingface.co