Tencent launches Hy4 open-weight model with vision and finance capabilities
The new flagship model features expanded training scales and specialized skills for 3D design, game development, and professional financial analysis.
The short version
- Hy4 features 770 billion parameters, a significant increase from the 300 billion parameters found in Hy3.
- The model includes new vision capabilities for analyzing dense graphics, documents, and multi-step visual reasoning.
- In a blind test of 579 financial tasks, Hy4 ranked first among open-weight models for overall performance.
- Tencent is offering a 50% discount on Hy4 via Tencent Cloud TokenHub until December 31, 2026.
- The model is compatible with Computer Use, allowing it to operate via command lines and graphical interfaces.
Why it matters
For developers and businesses, Hy4 provides a high-performing open-weight alternative that can handle complex professional tasks like 3D scene building and financial research. Its integration into existing tools like WorkBuddy shows how large models are moving from general chat to specific office productivity.
A new flagship for open weights
Tencent recently introduced Hy4, the latest generation of its large language models. This release marks a major milestone for the company as it moves to compete with the world's most advanced open-weight models. Open-weight models are AI systems where the underlying parameters are available for the public to download and run on their own hardware.
The new model represents a significant jump in scale compared to its predecessor, Hy3. While Hy3 contained fewer than 300 billion parameters, Hy4 has been scaled up to 770 billion parameters. Tencent plans to release the model's weights and technical specifications shortly to the public.
To achieve these results, the Tencent Hy team expanded the amount of information used and the magnitude of the training process for both the initial and subsequent learning phases. These improvements have led to better instruction tracking and an enhanced ability to understand long contexts. The model is also designed to handle long-range agentive capabilities, which allow it to perform multi-step tasks over extended periods.
Vision and professional tools
One of the most notable additions to Hy4 is its new vision capabilities. Unlike previous versions, Hy4 can now perform visual understanding, reasoning, and tool usage. This means the model can interpret dense graphics, analyze complex documents, and cross-reference information across multiple images simultaneously.
The model has been specifically trained with professional software like CAD applications and Blender. This specialized training allows Hy4 to perform sophisticated visual reasoning rather than just basic image recognition. It can now perform multi-step analyses and deductions based on visual information provided to it by the user.
Because of these vision skills, Hy4 is compatible with Computer Use for the first time. This allows the model to operate through both command line interfaces and graphical user interfaces. It can handle a wide range of professional and everyday tasks by interacting with software in a way that mimics human computer use.
Performance in finance and gaming
Tencent conducted extensive testing to see how Hy4 performs in specialized professional fields. In the financial sector, the model was tested on 579 real financial tasks, including fund research and stock analysis. In a blind evaluation, Hy4 ranked first among all evaluated open-weight models for overall performance.
The evaluation involved 153 experts from top-tier financial institutions who were unaffiliated with Tencent. Approximately seven out of ten of these experts had more than five years of experience in the field. These results suggest that the model is capable of handling high-level quantitative strategies and complex financial data analysis.
The model also shows significant progress in video game design and frontend development. It was trained with developers from Tencent’s IEG to improve its ability to implement game mechanics and handle assets. The model has achieved deep adaptation to major engines like Unity, Godot, and UE, showing improvements of 19.1% and 9.8% in specific game development evaluations compared to the Hy4 preview.
Integration and accessibility
Hy4 is being integrated across Tencent’s entire product ecosystem, specifically enhancing the user experience of WorkBuddy. WorkBuddy has already become China’s largest office productivity AI agent based on its user base. It serves 60% of China’s top 100 companies and 50% of the top 500 companies in over 50 different sectors.
The company is maintaining a cost-efficient approach by keeping the pricing of Hy4 the same as the Hy4 preview. The costs are set at $0.834 per million tokens for input and $2.501 per million tokens for output. To encourage adoption, Tencent is offering a 50% discount on the listed price on Tencent Cloud TokenHub until the end of 2026.
Developers can also access the model through several third-party platforms, including OpenRouter, Hermes, and OpenCode. Additionally, a version called Hy4 Flash will be launched soon. This version is intended to offer the best price-performance ratio by significantly reducing the cost per task while retaining the core capabilities of the flagship model.
What to watch
- The release of the Hy4 Flash model and its specific cost-per-task metrics.
- How the model performs in real-world 3D scene building and web development tasks.
- The release of Hy5 and how many parameters it will contain.
Try it
Check if your preferred AI platform, such as OpenRouter, has added the Tencent Hy4 model to its available API list.
Sources and links
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by elciudadano.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
