Recommended Articles
Tech & AI

Tencent Launches Hy4, a 770-Billion-Parameter AI Model That Can See and Operate Computers

By Angeled · November 5, 2026 · Tech & AI
Tencent Launches Hy4, a 770-Billion-Parameter AI Model That Can See and Operate Computers
Tencent Launches Hy4, a 770-Billion-Parameter AI Model That Can See and Operate Computers

Tencent released the full version of Hy4, its flagship large language model, on November 4, 2026. The release is the successor to the Hy4 preview that Tencent open-sourced in late August 2026, and it adds two capabilities that are new to the Hunyuan model family: vision, meaning the model can read images and charts, and computer use, meaning it can operate a computer through its graphical interface and command line, including 3D and CAD tools. The model is available now through Tencent Cloud TokenHub, OpenRouter, and Tencent's own applications, including Yuanbao, WeCom, and ima, as well as its WorkBuddy and CodeBuddy coding tools.

The scale is large. Hy4 is a mixture-of-experts model with 770 billion total parameters, of which 49 billion are active for any given token, and a context window of about one million tokens. That is a sharp jump from its predecessor Hy3, which had 295 billion total parameters. Mixture-of-experts is an architecture in which a model contains many specialist sub-networks, called experts, and only a small subset is switched on for each token. It is how labs build very large models whose cost per query stays manageable, because most of the parameters sit idle on any given request.

What Is New in the Full Release

The preview released in August was text-only. The November release adds fine-grained image processing, complex chart analysis, and mathematical visual reasoning, plus computer operation that combines perception, deep reasoning, and tool invocation. In practice, that means Hy4 can look at a screenshot or a document, understand what is on screen, and then take actions — clicking, typing, running commands — to carry out a task. That combination is the foundation of the AI agent, a system that does work rather than only answering questions.

Tencent says Hy4 now ranks among the strongest open-source models for the first time, a claim that should be treated as marketing until independent testing confirms it. In its own benchmark appendix, Tencent reports Hy4 winning 27 of 34 benchmarks against DeepSeek V4.1 Flash. Those numbers were compiled by Tencent itself, and independent evaluations have so far only tested the August preview, not the final release. Tencent also ran a finance evaluation with 213 finance professionals across 579 real-world tasks, in which it says Hy4 ranked first among open-weight models with a score of 3.04, ahead of DeepSeek V4.1 Flash at 2.97.

The Open-Weights Question

Here the story needs care, because the word "open" is doing more work than the facts currently support. The August preview shipped with its weights on Hugging Face on day one, under the Apache 2.0 license. The full November release does not. Tencent says it plans to release the model weights and a comprehensive system card "in the coming days," and it has lined up support from the vLLM and SGLang inference communities, but it has given no date. As of November 4, Tencent's Hugging Face account still listed only the preview, and OpenRouter showed Tencent Cloud as the model's only provider.

Until the weights are published, Hy4 cannot be downloaded, run, fine-tuned, or inspected by anyone outside Tencent. What is available today is a hosted service, not an open model. The licence for the full release has also not been stated. Founders and developers who plan around Hy4 should treat the open release as a promise, not a download, and build on the API in the meantime.

Pricing and Availability

API pricing is unchanged from the preview: 6 yuan per million input tokens and 18 yuan per million output tokens, roughly $0.83 and $2.50 at current rates, with cached input tokens at 0.3 yuan per million. Tencent Cloud TokenHub is offering a 50 percent discount through December 31, 2026, and usage in WorkBuddy and CodeBuddy is free from November 4 to 15. Tencent also says a lighter, cheaper Hy4 Flash version focused on speed and cost is coming soon, though it has given no specifications, price, or date.

Why It Matters for Founders and Investors

Hy4 is part of a pattern that now defines the frontier AI market: Chinese labs are shipping very large, very capable models at prices that undercut Western closed models, and they are doing it at a pace of weeks, not quarters. The same week brought Moonshot AI's open weights for Kimi K3.1 Pro, another multi-trillion-parameter model. For startups, that means the cost of intelligence keeps falling and the capabilities available by API keep expanding — vision, computer use, and million-token context are becoming commodities rather than differentiators.

The strategic question for founders is what to build on top. When the model layer is cheap, open, and fast-moving, the durable value sits in the workflow, the data, and the distribution — the things a model cannot commoditize. The open-weights delay is also a reminder: a model you cannot download is a dependency on a vendor's roadmap, however open its branding sounds.

Angeled.com connects founders with investors.

Sources: Tencent Hunyuan release announcement, November 4, 2026; Derrington Collaborative AI briefing, November 4, 2026; MacCus analysis, November 4, 2026; OpenRouter model listing, November 2026.

Read Next