Tencent has launched Hy4, the latest version of its Hunyuan artificial intelligence model family, adding visual understanding and computer-use capabilities for the first time. The model can analyse images, interpret complex charts, interact with command-line tools and operate graphical software, expanding its capabilities beyond conventional text-based AI assistance.
Hy4 is available through Tencent Cloud TokenHub and third-party platforms including OpenRouter. Tencent plans to release the model weights and a detailed system card in the coming days, but they are not yet publicly available. The launch strengthens Tencent’s position in the competition among Chinese AI developers, including DeepSeek, Moonshot AI and Z.ai, to build models capable of handling complex real-world tasks.
Hy4 Adds Vision and Computer Use
The biggest development in Hy4 is the addition of two capabilities that were not available in the earlier Hy4 preview: visual understanding and computer use.
With vision, the model can process images, examine detailed visual information, interpret complex charts and perform mathematical reasoning involving visual inputs. This makes it more useful for tasks involving documents, screenshots, diagrams and other image-based information.
Computer use allows Hy4 to interact with both command-line interfaces (CLI) and graphical user interfaces (GUI). In practical terms, an AI agent can use these capabilities to operate software rather than simply provide instructions for a human to follow.
Tencent said it has trained the model to work with professional applications such as Blender and computer-aided design (CAD) software. It has also improved its ability to handle complex, long-running tasks involving multiple steps.
The company has worked on integrating the model with development environments and game engines, including Unreal Engine, Unity and Godot. These capabilities could help developers automate parts of software development, build interactive interfaces and manage more complicated workflows.
Tencent Claims Improvements Across AI Benchmarks
Tencent says Hy4 delivers substantial improvements over its predecessor, Hy4 preview, across several evaluations.
In game development benchmarks, the company reported a 19.1% improvement in GameDev and a 9.8% improvement in GameLogic compared with the preview version.
The model has also received upgrades for financial analysis. Tencent said its evaluation involved 213 financial professionals assessing 579 real-world financial tasks. Hy4 achieved an overall score of 3.04, ranking ahead of the competing models included in the test.
| AI model | Score in Tencent’s financial evaluation |
|---|---|
| Tencent Hy4 | 3.04 |
| DeepSeek-V4.1-Flash-1017 | 2.97 |
| Kimi K3.1 Pro | 2.68 |
| GLM-5.3 | 2.68 |
The evaluation included professionals from external financial institutions, according to Tencent. However, the results were reported by the company and should not be treated as independently verified performance rankings.
Tencent also highlighted improvements in information retrieval and factual accuracy. In its internal evaluations, the model’s hallucination rate declined from 12.95% in Hy4 preview to 7.2% in the new release. Errors caused by unsupported assumptions fell from 23.1% to 5.6%.
These improvements are particularly relevant for applications such as financial research, business analysis and document processing, where incorrect information can have significant consequences.
AI Agents Move Beyond Chatbots
Hy4’s computer-use capabilities reflect a broader shift in the AI industry towards agent-based systems.
Traditional chatbots primarily respond to questions, generate text and assist with coding. AI agents are designed to take additional steps, such as opening applications, interacting with software, retrieving information and completing tasks.
For example, an AI assistant with computer-use capabilities could potentially inspect a design in Blender, make changes based on instructions and review the resulting output. In an office environment, similar capabilities could help automate document preparation, data analysis and repetitive administrative work.
Tencent is integrating Hy4 into WorkBuddy, its workplace productivity platform. The company says the model can work with information across chats, emails, documents, calendars and meetings through its integration with WeCom.
It can help identify pending tasks, track project developments, flag upcoming meetings and highlight approaching deadlines.
Tencent reported an average performance score of 83.18 and a pass rate of 96.15% for Hy4 across its WorkBuddy evaluation benchmarks. The company said the model ranked first in nine of 12 tested scenarios in mainland China.
These results are company-reported evaluations, but they demonstrate the direction Tencent is pursuing: moving AI from conversational assistance towards software that can participate directly in everyday work.
Hy4 Available on Tencent Cloud and OpenRouter
Tencent has made Hy4 available through its own cloud infrastructure and selected third-party platforms.
Users can access the model through:
- Tencent Cloud TokenHub
- OpenRouter
- CodeBuddy
- WorkBuddy
- Yuanbao
- ima
- WeCom’s AI Assistant
Its availability through OpenRouter is particularly important for developers who want to access models from different AI companies through a common API platform.
Tencent has kept the official API pricing broadly unchanged from the Hy4 preview.
| API usage | Tencent’s approximate pricing per million tokens |
|---|---|
| Input tokens | $0.90 |
| Output tokens | $2.70 |
| Cached input tokens | $0.045 |
OpenRouter’s model listing showed discounted rates at the time of reporting, with input tokens priced at $0.417 and output tokens at $1.251 per million tokens. Pricing and promotional offers can vary by provider and platform.
Tencent also announced a 50% discount on Hy4 through Tencent Cloud TokenHub until December 31, 2026. WorkBuddy and CodeBuddy were made available free of charge from November 4 to November 15.
The company is also preparing a smaller, more cost-efficient version called Hy4 Flash for users who prioritise lower costs and faster processing.
Open Weights Are Coming, but Not Available Yet
Despite describing Hy4 as part of its open-source AI strategy, Tencent has not released the new model’s weights alongside its launch.
The company said it plans to publish the weights and a comprehensive system card in the coming days. It is also working with the vLLM and SGLang communities to support the model’s deployment ecosystem.
The distinction matters because API access and open-weight availability are not the same thing.
Developers can currently use Hy4 through hosted services, but they cannot yet download the final model and independently deploy it on their own infrastructure. Publishing the weights would allow researchers and businesses to examine, customise and potentially fine-tune the model, subject to its eventual licence and technical requirements.
The earlier Hy4 preview was released as an open model in August 2026, with 770 billion total parameters, 49 billion active parameters and a context window exceeding one million tokens. Those figures describe the preview architecture and should not automatically be treated as independently confirmed specifications for every detail of the final release.
Tencent has not yet provided a specific public release date or licence for the final Hy4 weights.
Tencent Highlights Cybersecurity and Biosafety Measures
The release also comes with a focus on AI safety, particularly as models gain stronger coding, research and computer-use capabilities.
Tencent said Hy4 was trained using defensive cybersecurity material, including vulnerability analysis and remediation. It said offensive exploitation material was excluded from the training dataset.
The company’s Xuanwu Lab and Keen Security Lab also conducted security assessments covering areas such as cryptographic reverse engineering, binary analysis and vulnerability exploitation.
For biological safety, Tencent said internal expert reviews found that Hy4 has strong foundational biological knowledge but more limited capabilities in open-ended experimental design compared with frontier models.
The company plans to publish a system card alongside the model weights, providing additional information about its technical characteristics and safety evaluations.
The Bigger Picture
Tencent’s Hy4 launch highlights how competition in the AI industry is shifting from models that primarily generate text towards systems that can understand visual information and interact with software.
Vision and computer use are increasingly important capabilities for AI agents because they allow models to work with the same applications and interfaces used by people. This could create opportunities in software development, financial research, office productivity, design and scientific work.
The delayed release of the final model weights also highlights an important distinction in the open AI ecosystem. Making a model available through an API gives developers access to its capabilities, but releasing its weights provides greater flexibility for independent deployment and customisation. Tencent’s upcoming release will therefore be important for determining how accessible Hy4 becomes to the wider developer community.
Looking Ahead
Tencent is expected to release Hy4’s model weights and system card in the coming days, although it has not announced an exact date. The release will help developers understand the model’s technical requirements, safety characteristics and licensing conditions. Hy4 Flash could also expand access by offering a more affordable alternative for applications that do not require the full model’s capabilities.
The larger competition will depend on whether Hy4 can translate its reported benchmark improvements into reliable performance in everyday applications. With DeepSeek, Moonshot AI and other Chinese AI developers continuing to advance their models, Tencent is positioning Hy4 as a contender in the growing market for open-weight AI and autonomous agents. Its ability to operate software, interpret images and complete longer tasks could make it a significant development in the next phase of AI competition.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
