Executive summary
On November 4, 2026, Tencent released Hy4, a 770-billion-parameter flagship model that adds visual understanding and computer-use capabilities for the first time. It ranked first among open-source models in a financial blind evaluation with a score of 3.04, and Tencent plans to publish the model weights and system card in the coming days.
What happened
Tencent (00700.HK) announced the general availability of Tencent Hy4 at an event in Shenzhen on November 4, 2026. In its official press release, the company said the new model marks a significant leap in performance and capability and now ranks among the world's leading open-source AI models. Tencent Hy plans to release the model weights and a system card in the coming days.
Hy4 is the production version of the model previously distributed as Hy4 Preview. Tencent scaled up both pre-training and post-training for the full release. The disclosed parameter count is 770 billion (770B). Compared with the preview, Hy4 improves instruction following, long-context comprehension and long-horizon agentic tasks, and adds two capabilities for the first time: visual understanding and computer use.
On launch day, Hy4 became available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, the WeCom AI assistant (Dayuan) and ima. Its API is also offered through third-party platforms including OpenRouter, Hermes and OpenCode. Tencent said integration work is underway for QQ, Honor of Kings, Game of Peace, WeRead and Tencent News.
Key facts
| Item | Detail |
|---|---|
| Release date | November 4, 2026 |
| Company | Tencent (00700.HK) |
| Model | Tencent Hy4, general release |
| Parameters | 770B |
| New capabilities | Visual understanding; computer use |
| Office benchmark | WorkBuddy average score 83.18, pass rate 96.15%, first in 9 of 12 mainland China scenarios |
| Financial blind test | 213 finance professionals, 579 real tasks, Hy4 scored 3.04 and ranked first |
| Agent benchmark | Internal iRaB score up 12.03 percentage points over Hy4 Preview |
| Hallucination rate | Down from 12.95% in Hy4 Preview to 7.2% |
| Unsupported-assumption errors | Down from 23.1% to 5.6% |
| Game development | GameDev benchmark +19.1%, GameLogic +9.8% |
| API pricing | About S$0.90 per million input tokens, US$2.70 per million output tokens, US$0.045 per million cached input tokens |
| Promotions | WorkBuddy and CodeBuddy free from November 4 to 15; TokenHub 50% off through December 31, 2026 |
| Open-source plan | Weights and system card to be released in the coming days |
Vision: from reading images to multi-image reasoning
Hy4 gains visual understanding for the first time. Tencent says the capability covers fine-grained image processing, complex chart analysis, mathematical visual reasoning and reasoning across multiple images in one task. For office work, the model can read charts, scanned documents and multi-page files directly and combine visual information from several sources, rather than converting images to text before processing.
Tencent frames the capability as part of a co-design effort rather than an add-on image encoder. After identifying a chart, the model can call calculation or retrieval tools to complete the next step. The goal is to let visual input participate in the full task chain: read, reason, call tools and produce a result.
Computer use: operating CLI and GUI
Hy4 introduces computer-use capabilities for the first time. It can interact with command-line interfaces and graphical user interfaces to perform work and everyday tasks across applications. Tencent says the model received specialized training on professional software such as Blender and CAD, which strengthened its ability to plan and execute long-horizon tasks.
In multi-application workflows, the model has to decide which application to open first, how to pass intermediate results between programs and when to verify output. Tencent describes these long-horizon agentic tasks as a focus of the post-training stage and as the area where the full release differs most from the preview.
Office, finance, gaming and research
Office collaboration. On Tencent WorkBuddy, Hy4 scored an average of 83.18 with a pass rate of 96.15%, ranking first in nine of twelve scenarios in mainland China. Tencent says the model balances efficiency and quality in visual creative work, 3D scene construction and web development, and improves document presentation, stability and usability in human-AI collaboration. Through WorkBuddy's integration with WeCom, Hy4 can connect chat, email, documents, calendars and meeting data, track multiple projects, identify outstanding tasks and flag upcoming meetings and deadlines.
Financial analysis. The Tencent Hy team and Tencent Financial Technology ran a blind evaluation with 213 finance professionals, 153 of them from leading external institutions and about 70% with more than five years of industry experience. The evaluation covered 579 real financial tasks. Hy4 scored 3.04, followed by DeepSeek-V4.1-Flash-1017 at 2.97, and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. Hy4 ranked first among the open-source models evaluated.
Game development. After working with professional developers at Tencent Interactive Entertainment Group (IEG), Hy4 integrated deeply with Unreal Engine, Unity and Godot, supporting complex game logic and multiplayer implementations. Compared with Hy4 Preview, the GameDev benchmark improved 19.1% and GameLogic 9.8%.
Scientific research. On TenureTrail, which evaluates 123 executable tasks across seven disciplines, Hy4 performed strongly in mathematical construction, physical modeling, environmental and earth sciences, systems engineering and socio-economic decision-making. Tencent also open-sourced two foundation models built with Hy4 and the Hyra 2.0 harness: HyCell for single-cell representations and HyPDE for physics fields.
Open source, pricing and safety constraints
Hy4 keeps the same API pricing as Hy4 Preview: about S$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens. Tencent presents the unchanged price alongside materially better performance as part of its "practical AI" positioning. A lower-cost Hy4 Flash model is scheduled for release soon.
On open source, Tencent plans to publish Hy4's weights and system card in the coming days. The company says cybersecurity and biosafety were assessed before release. Training data focused on defensive security, including proof-of-concept development and vulnerability remediation, while material related to exploitation was explicitly excluded. Tencent's Xuanwu Lab and Keen Security Lab evaluated cryptographic reverse engineering, binary deobfuscation and vulnerability exploitation, and found no significant concerns about Hy4's offensive cybersecurity capabilities in the areas assessed. On biosafety, Tencent says Hy4 has strong foundational biological knowledge but more limited open-ended experimental design capability than frontier models.
Context and implications
Hy4 pushes open-source competition into a phase where several capabilities are added at once. Earlier open-source releases competed mainly on text reasoning and code generation. Hy4 adds vision, computer use and long-horizon agency in a single release and targets office, finance, gaming and research workflows, which shows that vendors are competing on real work rather than on a single benchmark.
For developers, 770B parameters make local deployment expensive. In the near term, the practical route is an API through Tencent Cloud TokenHub or third-party platforms such as OpenRouter. The free window and the 50% TokenHub discount indicate that Tencent wants to expand call volume quickly through price and distribution.
The results also need to be read with the right scope. Most benchmarks in the release were produced internally or commissioned by Tencent. The financial blind test did include 153 external experts, but the task list, scoring rubric and error margins were not fully disclosed. Until independent replication, the score differences should not be treated as a direct measure of business performance.
Official sources
- Tencent official press release (PR Newswire): Tencent Launches Hy4: Joining the Ranks of the World's Leading Open-Source AI Models
- AAP reprint: Tencent Launches Hy4
- TechNode: Tencent launches Hy4 with vision and computer-use capabilities
- Tencent Hy4 product page: Hy4 Research