--- title: "Tencent Unveils Hy4: Open-Weights 770B AI Model with Visual Perception" description: "Tencent has introduced the open 770B Hy4 AI model with visual perception support. The system is integrated into the WorkBuddy platform and outperforms several competitors, though it trails closed flagships in complex tasks." date: 2026-11-04T23:48:05.497Z lang: en url: https://xab.info/en/posts/tencent-hy4-open-770b-ai-model tags: [tencent, hy4, ai-model, open-source, artificial-intelligence, multimodal-ai] author: "Alexander Buran" publisher: "XAB.info" --- # Tencent Unveils Hy4: Open-Weights 770B AI Model with Visual Perception ![Graphical diagram of the Tencent Hy4 AI model with geometric constructions and letter markings](https://xab.info/media/2026/11/04/tencent-hy4-otkrytaya-ii-model-770-mlrd-parametrov/tencent-hy4-otkrytaya-ii-model-770-mlrd-parametrov-1.webp) ## Key Points - Tencent Hy4 is an open 770B parameter model, 2.6 times larger than Hy3 - Key innovation is visual perception: images, charts, mathematical reasoning - Paired with WorkBuddy, the model scored 83.18 points with a 96.15% success rate - Hy4 outperformed Z.ai GLM-5.3 and DeepSeek-V4.1-Flash-1017, but trails closed flagships in complex tasks - Hy5 (larger) and a lightweight Hy4 Flash are coming; cyber capabilities were restricted during training Chinese tech giant Tencent Holdings has unveiled its flagship artificial intelligence model, Hy4, released with open weights and focused on advanced AI agent capabilities. The new development is positioned as the foundation for the WorkBuddy enterprise platform and, according to the developers, makes the company's line of AI solutions "one of the most powerful among currently available open-weights models." ### Visual Perception as a Key Innovation The most important distinction between Hy4 and its predecessors is the support for visual perception: the model is capable of processing images, analyzing complex charts, and performing mathematical reasoning based on visual data. This expands the system's application scenarios beyond text-based tasks and brings it closer to closed-segment multimodal solutions. At the same time, Tencent honestly notes that Hy4 still lags behind leading closed systems, especially in solving complex tasks such as mathematical research, writing code within long-term projects, self-improvement, and advanced visual analysis, including answering questions about images. ### Scale: 770 Billion Parameters and Plans for Hy5 The size of Tencent Hy4 is 770 billion parameters — 2.6 times larger than its predecessor Hy3 with 295 billion parameters. The developer promised that the next generation of the model, Hy5, will be even larger, and a lightweight Hy4 Flash variant will appear in the near future to significantly reduce task execution costs. This approach allows the company to cover various segments: from heavy research scenarios to economical industrial use. ### Integration with WorkBuddy and Benchmark Rankings Tencent has integrated the new model into the WorkBuddy AI agent platform: according to internal estimates, when paired with this tool, Hy4 showed an average score of 83.18 points with a 96.15% success rate for solutions. The preliminary version of Hy4, introduced in late August, currently ranks 23rd in the global WebDev ranking on the Code Arena platform — a benchmark that evaluates large language models in real-time based on practical programming tasks. ### Supremacy Over Predecessor and Competitors A significant increase in training data volume and model scale helped Hy4 outperform its previous version in 12 task categories, including autonomous action execution, code writing, and workflow automation. Compared to competitors, the model bypassed Z.ai GLM-5.3 and DeepSeek-V4.1-Flash-1017, performing on par with Moonshot Kimi K3.1 Pro. Thus, Hy4 solidifies its position in the upper tier of open models, although a gap remains before achieving absolute leadership among closed systems. ### Cybersecurity Limitations During the training process, Tencent deliberately restricted capabilities related to hacking activity: materials associated with exploit development, reverse engineering, and penetration testing were excluded from the training dataset. The company explains this by the need to "avoid enhancing capabilities for conducting offensive cyberattacks." This approach reflects the growing practice of responsible AI model training, where developers embed restrictions on potentially dangerous use cases during the data preparation stage. ## Fact-Check & Sources - [Китайская Tencent бросила вызов DeepSeek и Kimi — открытая ИИ-модель Hy4 умеет видеть и рассуждать window-new](https://3dnews.ru/1150848/kitayskaya-tencent-brosila-vizov-deepseek-i-kimi-otkritaya-iimodel-hy4-umeet-videt-i-rassugdat) - Primary source: model parameters (770B), visual perception, WorkBuddy (83.18 points, 96.15%), 23rd place in Code Arena WebDev, comparison with Z.ai GLM-5.3, DeepSeek-V4.1-Flash-1017, Moonshot Kimi K3.1 Pro, cybersecurity restrictions. - [Foreign Affairs: DeepSeek бросила вызов всему американскому рынку ИИ](https://tass.ru/ekonomika/23269443) - Подтверждено по источнику tass.ru - [Foreign Affairs: DeepSeek бросила вызов всему американскому рынку ИИ](https://news.am/ru/news/869121) - Подтверждено по источнику news.am - [DeepSeek: вызов гигантам или временная сенсация? Чем отличается от конкурентов и ...](https://www.ixbt.com/live/sw/deepseek-vyzov-gigantam-ili-vremennaya-sensaciya-chem-otlichaetsya-ot-konkurentov-i-konkurentosposoben-li-na-samom-dele.html) - Подтверждено по источнику ixbt.com ## FAQ ### How many parameters does the Tencent Hy4 model have? Hy4 is 770 billion parameters in size, which is 2.6 times larger than the previous Hy3 model with 295 billion parameters. ### What is the main innovation in Hy4? Support for visual perception: the model processes images, analyzes complex charts, and performs mathematical reasoning based on visual data. ### Which models did Hy4 outperform in comparisons? Hy4 outperformed Z.ai GLM-5.3 and DeepSeek-V4.1-Flash-1017, performed on par with Moonshot Kimi K3.1 Pro, and beat its previous version across 12 task categories. ### What restrictions were introduced during Hy4 training? Materials on exploit development, reverse engineering, and penetration testing were excluded from the training set to prevent enhancing offensive cyberattack capabilities. ### What future model versions are expected? The developer promised a larger Hy5 model and a lightweight Hy4 Flash variant to reduce task execution costs.