Tencent Launches Hy4: Vision and Computer Use, Open Weights Coming
· Cortex Solutions
On Nov 4, Tencent released the full Hy4 with vision and computer use, same API prices, and open weights promised later. What changed and what it means for Saudi firms
On November 4, 2026, Chinese tech giant Tencent announced the full release of its flagship large language model Hy4, roughly two months after a preview version appeared in August. According to a TechNode report published the same day, the new release adds two capabilities the preview lacked: visual understanding and direct computer use. Unite.AI also covered the announcement on November 4, with more detail on evaluations and safety measures.
What is new in Hy4?
According to TechNode, the model can now interpret charts and documents, reason across multiple images, and operate command-line tools and graphical interfaces. Tencent says it applied targeted training on professional software such as Blender and CAD tools, as well as on tasks that span several applications.
The preview, open-sourced in August 2026, was a text-only mixture-of-experts model with 770 billion parameters, 49 billion of them active per token, and a context window of about one million tokens, as reported by Unite.AI and Channel Insider. TechNode puts the full release at 770 billion parameters as well.
Unite.AI adds that Tencent substantially scaled both pre-training and post-training, with gains in instruction following, long-context understanding, long-horizon tasks and deep research.
The numbers, as Tencent reports them
Unite.AI relayed several results from Tencent's own evaluations, which have not yet been independently verified:
- In a finance evaluation with 213 finance professionals across 579 tasks, Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 (2.97) and Kimi K3.1 Pro and GLM-5.3 (2.68 each), ranking first among open-weight models.
- The hallucination rate in internal search evaluations fell from 15.76% in the preview to 5.45%.
- In game development tasks, performance improved by 19.1% on GameDev and 9.8% on GameLogic compared with the preview.
At the same time, according to Unite.AI and analyst Pasquale Pillitteri, Tencent acknowledges that the model still trails the best proprietary models in research-level math, long-horizon coding and advanced visual question answering.
Pricing and availability
TechNode reports that API pricing is unchanged from the preview: $0.834 per million input tokens and $2.501 per million output tokens, with a 50% discount on Tencent Cloud TokenHub through December 31. Unite.AI adds that cached input costs $0.042 per million tokens.
The model is available through Tencent Cloud and OpenRouter, and inside Tencent products such as CodeBuddy, WorkBuddy, Yuanbao and ima. TechNode notes that WorkBuddy and CodeBuddy users get a limited free usage quota from November 4 to November 15.
The open weights of the full release have not been published yet. Tencent says it will release them together with a system card shortly, without giving a date. It also announced a lighter, lower-cost Hy4 Flash model for the near future.
Safety: defensive cybersecurity data only
One notable point in the announcement, according to Unite.AI and Pillitteri, is that Tencent limited cybersecurity training data to defensive content, such as proof-of-concept generation and vulnerability patching, and filtered out exploit development, reverse engineering and penetration-testing material. Tencent Xuanwu Lab and Tencent Keen Security Lab, both part of Tencent, ran red-team audits and concluded that the model does not raise risk on those tasks. Pillitteri points out that this conclusion was not accompanied by figures.
What does this mean for Saudi businesses?
This is Cortex's editorial reading, not part of Tencent's announcement:
- Competition among open-weight models is intensifying, which lowers the cost of running AI. A price below one dollar per million input tokens makes automating everyday tasks viable even for small and medium businesses.
- Computer use means an AI agent can work on existing systems through their interfaces, not only through APIs. That helps with legacy systems that have no API, but it requires controlled permissions and human review.
- Once the weights are published, it becomes possible to run a strong model on servers inside the Kingdom, which matters for organisations handling sensitive data under personal data protection requirements.
- The published figures come from the company's own evaluations. It is best to test any model on your actual documents and tasks in Arabic before relying on it.
If you are considering AI agents for your business, from choosing the right model to connecting it to your systems and setting permission controls, learn about Cortex's AI automation service: https://cortex.sa/services/ai-automation
Sources: TechNode (November 4, 2026), Unite.AI (November 4, 2026), Pasquale Pillitteri (November 4, 2026), Channel Insider (August 31, 2026, on the preview).