NEWS

Tencent releases Hy4 weights, multimodal model with 1 million token context

Tencent releases Hy4 weights, multimodal model with 1 million token context
Image: Redação iMasters
News
Reading
4 min
In short

China's Tencent released the weights of Hy4, its flagship multimodal model, under the Apache 2.0 license. It has 770 billion parameters, a context window of around one million tokens, and support for text, image, and video.

What Tencent released

Tencent released the weights of Hy4, its most advanced AI model, this Friday (November 20), allowing any developer to download and run the model on their own infrastructure. Hy4 had been announced on November 4, but until now it was only accessible via Tencent's own API. The model is now available under the Apache 2.0 license, which allows commercial use and modification without royalty restrictions, according to TechNode.

The flagship release follows the earlier opening of Hy4 flash, a smaller version from the same family. The strategy of releasing the compact model first and then the full flagship model later has become common among Chinese labs that bet on open weights as a way to gain adoption outside China.

Inside the model: a 770-billion-parameter MoE architecture

Hy4 uses a mixture-of-experts (MoE) architecture: the model has 770 billion parameters in total, but activates only 49 billion to process each token. This is the same logic behind other open MoE models circulating in the market today: the model's total size is enormous, but the computational cost per inference stays closer to that of a much smaller model, because only a fraction of the parameters come into play on each call.

The context window is around one million tokens, enough to process entire code repositories, lengthy documents, or hours of transcription in a single call. The model is multimodal: it understands text, image, and video. According to Tencent, the Hy4 released now improves on the preview version disclosed in November on four fronts: instruction following, long-context understanding, extended tasks, and research.

What it's for in practice

Tencent positions Hy4 for workflows that go beyond conventional chat. The model was designed to handle graphical interfaces and command-line tools, as well as frontend development tasks, game development, and office task automation.

This brings Hy4 closer to a category of model designed for agents that operate directly on development environments, not just to answer text questions. With a one-million-token context, the model can in theory keep an entire software project in memory while executing a task end to end, something that models with smaller windows force developers to work around with truncation or RAG.

How to run Hy4 (and what it requires)

Tencent published deployment instructions for two serving tools already familiar to those who work with open-weight LLMs: vLLM and SGLang. It also provided resources for fine-tuning and a safety evaluation card, a document detailing the safety tests performed before launch.

A practical warning is worth noting: a 770-billion-parameter MoE model, even activating only 49 billion per token, still needs to load the entire model into GPU memory to work. This means that, in practice, running the full Hy4 requires a cluster with multiple high-capacity GPUs, not a single developer machine. Those who want to test it without that investment tend to turn to Hy4 flash, the smaller version already opened earlier, or to cloud providers that offer the model hosted.

What changes for developers in Brazil

In short: one more flagship model becomes an option outside the closed API, and that weighs on the choice of anyone building an AI stack today. The Apache 2.0 license is the detail that matters most: it allows commercial use and modification of the model without requiring licensing or royalties, unlike the more restrictive licenses some labs apply to their own open weights.

For Brazilian teams weighing whether to run models via API (Claude, GPT, Gemini) or host something of their own, Hy4 enters as one more piece on that board. The advantage of running it locally is cost predictability and data that never leaves the team's own environment; the disadvantage is the infrastructure barrier a model this size imposes. This brings Hy4 closer to companies with a solid cloud GPU budget, or to those who already operate their own inference infrastructure, than to small squads looking to run a model on a laptop.

What remains open

The source does not provide numerical benchmarks comparing Hy4 to competitors such as GPT, Claude, or other open Chinese models like DeepSeek and Qwen. Tencent claims improvements in instruction following, long context, extended tasks, and research compared to Hy4's own preview version, but without publishing numbers that allow direct comparison with models from other labs.

Also, the material released so far includes no details on estimated hosting costs, minimum GPU hardware requirements, or independent benchmarks run by third parties. That data should emerge in the coming weeks, as developers test the model outside Tencent's controlled environment.

Translated from the Brazilian Portuguese original · Read the original