Tencent has released the full version of Hy4, its latest flagship artificial intelligence model, adding two capabilities the preview version lacked: the ability to understand images, and the ability to operate a computer. Announced on 4 November 2026, Hy4 is the successor to the Hy4 Preview that arrived in August, and Tencent says the upgrade is a significant leap in performance that puts the model among the world’s leading open-source systems.
Hy4 is a mixture-of-experts system with 770 billion parameters in total, of which 49 billion are active for any given task, and it can take in around one million tokens of context at once — roughly the text of several long books in a single prompt. The preview already had the scale. What is new is what the model can perceive and do.
Vision arrives for the first time
For the first time in Tencent’s Hunyuan line, Hy4 ships with native visual understanding. According to Tencent, it can read fine-grained details in images, work through complex charts, and solve mathematical problems presented as pictures. That matters because much of real work is visual: a chart in a report, a diagram in a manual, a screenshot of an error message. A model that can only read text has to be told what the picture shows; Hy4 can look for itself.
Computer use: from answering to acting
The bigger shift is computer use. Hy4 can now interact with command-line interfaces and graphical user interfaces to carry out tasks, rather than merely describing how to do them. Tencent says the model received specialised training on professional software such as Blender, for 3D creation, and CAD tools, and that it is built to plan and execute long, multi-step jobs — the kind of work that involves opening applications, moving between them, and checking its own progress.
This is the direction the whole frontier is moving: from chatbots that answer questions to agents that do jobs. Anthropic’s Claude, OpenAI’s ChatGPT agent features and Google’s Gemini have all pushed in the same direction, and Hy4 is Tencent’s bid to stay in that race with an openly available model. Tencent’s own figures, from an internal blind evaluation of 579 real financial tasks judged by 213 professionals, put Hy4 first among open-weight models at 3.04, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68. Those are the company’s numbers, not independent results, and they should be read that way until outside labs test the full model.
What it costs, and the free window
Access is the other half of the announcement. Hy4 is available now through Tencent Cloud TokenHub, inside Tencent products including WorkBuddy, CodeBuddy, Yuanbao and the WeCom assistant Dayuan, and through third-party platforms such as OpenRouter. Despite the performance gains, Tencent says pricing is unchanged from the preview: around $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens.
There are two time-limited offers. WorkBuddy and CodeBuddy are free from 4 to 15 November, and Hy4 on Tencent Cloud TokenHub is half price through 31 December 2026. A cheaper and faster variant, Hy4 Flash, is promised soon but has no release date yet.
The catch: the weights are not out yet
Read the announcement carefully and one qualification stands out. The Hy4 Preview was open-sourced in August, but for this full version Tencent says only that it plans to release the model weights and a full system card in the coming days. Until those files actually appear, Hy4 is a hosted release: you can use it through Tencent’s services and partners, but you cannot yet download and run it yourself. Tencent’s announcement frames the model as open-source, and the weights may well follow within days, but the label is ahead of the files for now.
Tencent also describes safety work around the release. Hy4 was trained on defensive cybersecurity data — proof-of-concept development and vulnerability fixes — with offensive material about exploitation explicitly excluded, and the company says its security labs found no significant concerns about offensive capability in the areas they assessed. On biology, the company says the model has strong foundational knowledge but is more limited at open-ended experimental design than frontier rivals.
Why this matters
Hy4 matters for two reasons beyond Tencent. First, it narrows the gap between China’s open-weight models and the closed frontier: a model that can see screens and operate software, at open-model prices, is a practical tool rather than a research demo. Second, it raises the stakes on a question every lab now faces — what “open” means when the weights lag the announcement. If Tencent publishes the weights and system card as promised, Hy4 becomes one of the most capable downloadable models available. If the delay stretches, the release will be remembered as a hosted product launch wearing an open-source label. Either way, the next few days, when the files are due, are the real test of this announcement.
Related reading: our earlier explainer on Google’s Gemini 4 Argon, and our report on a new benchmark where even the best AI scored just 12% on real work. Tencent’s own announcement is the primary source for this article: Tencent launches Hy4.
One Comment