Tencent Launches 770 Billion Parameter Hy4, Bringing Vision and Computer Control to Its AI Tools

Hy4 moves from answering questions to handling work

Tencent released the formal version of its Hunyuan Hy4 artificial intelligence model at noon on November 4, 2026, adding visual understanding and computer operation to a system with 770 billion parameters. The launch puts the model inside workplace, coding and consumer products while Tencent prepares to release its downloadable model weights.

The company says Hy4 has entered the ranks of the strongest open models, with better instruction following, understanding of long documents and execution of tasks involving several steps. Its developers also acknowledge a substantial gap between Hy4 and the leading closed models, whose underlying weights remain controlled by their owners.

Tencent is making its case through workplace results rather than size alone. In a financial task evaluation organised with Tencent Financial Technology, known as FiT, 213 experts assessed 579 real tasks. Hy4 scored 3.04, ahead of the other open models included in that exercise.

The release also brings a practical change: Hy4 can interpret images and work through graphical computer interfaces as well as command lines. Tencent has trained it for specialist software including Blender and CAD tools, alongside tasks that move between applications. For users, the intended shift is from receiving instructions to having software carry out more of the work.

From the August preview to the November release

Hy4's development spans two distinct releases. The preview arrived on August 28 with 770 billion total parameters, 49 billion active parameters per token and a context window of about one million tokens. The November version adds capabilities and improvements while retaining the headline model size.

Tencent's Hunyuan team described the preview's purpose in its launch statement:

Today we're releasing Hy4 preview, our most capable model to date,

The company has since pushed the model further toward office work, software development, finance and research. Its release sequence also separates the already distributed preview from the final version, whose weights and system card are still scheduled for publication.

• February 2026: Tencent rebuilt Hunyuan's infrastructure, after which Tencent Cloud says the team averaged a major model update every two months.
• August 28: Hy4 preview launched with 770 billion total parameters and an Apache 2.0 licence.
• October 29: Dowson Tong, Tencent senior executive vice president and chief executive of its Cloud and Smart Industries Group, announced the approaching Hy4 and Hy4 flash releases at Tencent's Global Digital Ecosystem Summit.
• November 4: The formal Hy4 release went live across several Tencent products.
• November 4 to November 15: WorkBuddy and CodeBuddy are offering users a limited allowance of free access.
• December 31, 2026: Tencent Cloud TokenHub's advertised discount period ends.

What 770 billion parameters means

Parameters are the numerical settings a model learns during training. They help determine how it recognises patterns, follows instructions and generates answers. Hy4 has far more of them than the preceding Hy3 series, which had fewer than 300 billion.

The August preview uses a mixture of experts architecture. Rather than applying every part of the model to every token, it routes work to selected specialist networks. A token is a small unit of text, such as a word or part of a word.

About 49 billion parameters are active for each token in the preview, roughly 6.4% of its 770 billion total. Its architecture has 78 layers. Most contain 256 routed experts, with eight selected for each token, plus a shared expert. This selective approach reduces the computation needed for each step while keeping a much larger pool of learned information.

The preview's context window holds about one million tokens. Context is the material a model can work with at once, including instructions, documents and earlier exchanges. A large window is useful for projects that involve many files or lengthy records, although the quality of the model's reasoning still matters when it must connect details spread across them.

Running a model of this size requires substantial memory and computing infrastructure. Selecting fewer experts reduces calculation, but the full set of weights still needs storage and access. For most users, a cloud service or an application such as WorkBuddy will be the practical route to Hy4.

Vision and computer use expand the job list

Hy4 is the first Hunyuan flagship to gain visual understanding. Tencent says it can perform detailed image processing, analyse complex charts and solve mathematical problems presented visually. That gives it a way to use information that is difficult to capture in a plain text prompt.

Computer use adds another layer. Hy4 can combine command line operations with actions in graphical interfaces, allowing it to work across software rather than relying entirely on text replies or predefined connections to services. Tencent says it conducted specialist training for Blender, CAD software and extended tasks involving several applications.

An AI agent is a system that takes steps toward a goal, uses tools and adjusts what it does in response to results. Hy4's visual and computer capabilities broaden the kinds of steps such an agent can attempt. A chart, an application window or a design interface can become part of the task rather than something the user must describe manually.

Tencent also says the formal release handles lengthy tasks more effectively and uses its reasoning process more efficiently. Those improvements address limitations identified in the preview, including spending too long reasoning through difficult jobs and repeatedly checking its own work. For a user waiting on a completed document or software change, unnecessary checking adds time as well as processing costs.

WorkBuddy turns the model into an office product

Tencent has launched Hy4 through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, the WeCom intelligent assistant Dayuan and ima. Access through external platforms includes OpenRouter. Other Tencent services, including QQ, Honor of Kings, Peacekeeper Elite, WeChat Read and Tencent News, are continuing integration work.

WorkBuddy is central to the office strategy. Through its WeCom suite, Hy4 can connect work information across chats, email, documents, calendars and meetings. Bringing those records together can help it complete a task that would otherwise require a worker to search several places and assemble the relevant details.

Tencent's WorkBuddy evaluation gave Hy4 an average effectiveness score of 83.18 and a task pass rate of 96.15%. The company says it ranked first among domestic competitors in nine of 12 scenarios, or three quarters of the tested categories. The scoring scale and full task conditions were not detailed in the launch account.

Tencent also reports improvements in document appearance, stability and usability for WorkBuddy's feature that lets people and AI write together. The combination of model development and product development reflects Tong's stated focus on what users gain in a particular setting, rather than concentrating only on parameter counts and benchmark scores.

Financial tasks show progress and remaining errors

Finance received a dedicated training and evaluation effort. Tencent says it built thousands of workspaces and more than 10,000 high quality tasks based on real financial scenarios, then worked with FiT to organise the assessment by 213 experts.

Hy4's score of 3.04 across 579 financial tasks compared with 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3. The lead over DeepSeek was 0.07 points. In Xuntu Technology's internal IRAB evaluation, Hy4 improved by 12.03 percentage points from the preview.

Tencent's own tests also measured factual errors. After work on retrieving current information, combining sources and checking consistency, the hallucination rate fell from 12.95% in the preview to 7.2%. A hallucination is an answer that presents invented or incorrect information as fact. The reduction was 5.75 percentage points, or about 44% relative to the earlier rate.

Errors attributed to guessing fell from 23.1% to 5.6%, a reduction of 17.5 percentage points. Those remaining errors matter in finance, where a mistaken figure can carry through a spreadsheet and affect the final conclusion. Financial statements, formulas and citations still need human checking before they support an investment or business decision.

One documented WorkBuddy task illustrates the scale of work Hy4 can attempt. Asked to gather Tencent, Alibaba and Baidu financial reports from 2020 onward and compare their businesses from four professional perspectives, it took 41 minutes and 51 seconds. The output included a PDF of 12 pages, eight charts, financial data tables and an Excel workbook with 183 formulas that could be recalculated. A download log recorded addresses, status, file sizes and reasons for failures.

Benchmark gains need their own context

The August preview underwent a different Tencent evaluation: 163 internal experts assessed 203 engineering tasks without knowing which model produced each answer. Hy4 preview averaged 2.99 out of four, compared with 2.94 for Kimi K3 and 2.92 for GLM-5.3.

Against GLM-5.3, Hy4 preview won 46.8% of comparisons, tied 12.8% and lost 40.4%. Against Kimi K3, it won 51.2%, tied 7.9% and lost 40.9%. Those figures show a modest average advantage within that particular group of tasks, with competitors still winning many comparisons.

The engineering evaluation and November financial evaluation use different tasks, participants and rival model versions. Their scores describe separate exercises. Tencent's internal assessment provides detail on workplace performance, while independent testing remains a separate part of establishing how broadly the results carry across other jobs.

For the formal release, Tencent reports gains over the preview of 19.1% on GameDev and 9.8% on GameLogic. It says development work targeted complex frontend interactions, business logic, state management and connections between frontend and backend systems. In simpler terms, the goal is software that behaves correctly as users interact with it, rather than a page that merely looks finished.

Tencent also says Hy4 can combine scientific reasoning, programming and lengthy task execution. The company has released the HyCell model for single cell representation and the HyPDE model for physical fields, describing them as developed with assistance from Hy4 and Hyra2.0. Its stated preparation for recursive self improvement involves AI helping improve AI, while Hy4's immediate commercial push centres on products people can already use.

Prices stay unchanged as open release approaches

Hy4's listed API prices remain the same as the preview: 6 yuan per million input tokens, 18 yuan per million output tokens and 0.3 yuan per million cached input tokens. Input is the material sent to the model; output is what it generates. Cached input is previously processed material that a service can reuse more cheaply.

Tencent Cloud TokenHub is charging half the listed rates through December 31, 2026. That brings the corresponding prices to 3 yuan for input, 9 yuan for output and 0.15 yuan for cached input per million tokens during the promotion. A task using one million ordinary input tokens and one million output tokens would therefore cost 12 yuan at those discounted rates.

The August preview was distributed under Apache 2.0 through repositories including Hugging Face, ModelScope, GitCode and CNB. Tencent says the formal Hy4 weights and a system card will be released soon, but has given no exact publication date. A system card describes a model's capabilities, testing, limitations and safety considerations. Tencent says its release work will pay particular attention to cyber and biological security risks.

Hy4 flash is also planned for release soon as a cheaper option drawing on Hy4's capabilities. Tencent has not announced its launch date or prices. The company says the next generation, Hy5, will be larger, without confirming a release schedule.

Why Tencent is concentrating on office users

QuestMobile's September 2026 figures put monthly active users of AI office services at 142 million, up 50.6% from a year earlier. WorkBuddy had 32.353 million, compared with 9.77 million for Baidu's Dazi, 7.905 million for Qianwen Office and 2.049 million for Doubao Work.

WorkBuddy's audience was about 3.3 times Dazi's and equivalent to roughly 23% of the measured AI office audience. Those figures give Tencent a substantial position in a category where access to work records, useful documents and completed tasks can matter more to customers than the appeal of a general chatbot.

Tencent faces a different competitive position in consumer assistants. Yuanbao had 37.33 million monthly active users in September, compared with 419 million for Doubao, 158 million for Qianwen and 123 million for DeepSeek. WorkBuddy therefore gives Hy4 a product with a stronger position in its own category.

The commercial test is whether improved task completion persuades users and businesses to pay. Usage figures measure audience size, while Tencent's coming financial results will provide another opportunity to assess spending and revenue. Its third quarter results are expected in mid November. Hy4's launch connects the company's model investment to an existing office audience, with reliability, speed and cost now central to the customer experience.

Key Points

• Tencent released the formal Hy4 model on November 4, 2026, with 770 billion parameters and new visual understanding and computer operation capabilities.
• Hy4 is available in WorkBuddy, CodeBuddy, Yuanbao, TokenHub and other Tencent products.
• Tencent's financial task evaluation gave Hy4 3.04, ahead of the other open models tested.
• The company's own tests recorded a hallucination rate of 7.2%, down from 12.95% in the preview.
• Listed API prices remain 6 yuan per million input tokens and 18 yuan per million output tokens, with TokenHub charging half price through December 31.
• The formal model weights, system card and Hy4 flash are planned for release, with no exact dates announced.
• Tencent acknowledges that Hy4 still trails the leading closed models.