NederlandsTencent releases Hy4: an open AI model that understands images and operates a computer
China's Tencent has released Hy4, a large language model that understands images for the first time and can operate a computer on its own. According to Tencent, it makes things up far less often than the previous version. The model files will be freely available in the coming days.
The Chinese technology company Tencent released Hy4 on 4 November 2026, the latest large language model in its Hy family. For the first time, a model from this family can also understand images: it reads detailed photos, complex charts and maths problems that contain pictures. Hy4 can also operate a computer on its own, both through the command line (where you type commands) and through ordinary windows and buttons. For this, Tencent also trained the model on professional programs such as Blender and CAD software.
Tencent plans to release the model files, known as the weights, and a detailed description of the model in the coming days. Anyone will then be able to download and run the model. A first preview, Hy4 Preview, appeared at the end of August.
Making things up less often
A known problem of language models is that they sometimes state invented things with great confidence. This is called hallucinating. According to Tencent's internal tests, this happens in 7.2 per cent of cases with Hy4, compared with 12.95 per cent with Hy4 Preview. Errors caused by the model assuming something that has no basis fell from 23.1 to 5.6 per cent, according to Tencent.
Performance according to Tencent
Tencent mainly tested the model on work that people do in an office. For this, the company connected Hy4 to its own work program WorkBuddy, which is linked to the collaboration software WeCom. The model can then search chats, emails, documents, calendars and meetings and, for example, keep track of tasks and deadlines. According to Tencent, Hy4 passed 96 out of 100 tasks in this test, and was the best of the models on the market in China in nine of the twelve areas.
The model has also been specially trained for financial work. In a blind test, 213 people from the financial world, about 70 per cent of whom have worked in the field for more than five years, rated the answers to 579 real financial tasks. They did not know which model gave which answer. Hy4 scored 3.04 on average, just above DeepSeek-V4.1-Flash (2.97). The Chinese models Kimi K3.1 Pro and GLM-5.3 both scored 2.68. Tencent ran this test together with its own financial arm.
For people who make games, Hy4 works with the well-known game engines Unreal Engine, Unity and Godot. On two of its own tests for building games, the model improved by 19.1 and 9.8 per cent compared with the preview, according to Tencent.
Almost all the figures come from Tencent itself. Independent measurements often follow in the weeks after a launch like this, once other researchers can test the model themselves.
Safety and availability
During training, Tencent paid attention to possible misuse. The model mainly learned how to find and close security holes; according to Tencent, material about actively exploiting such holes was deliberately left out. Security experts from two of Tencent's own research groups, Xuanwu Lab and Keen Security Lab, tested the model and found no major concerns in the areas they examined. In biology, Hy4 has a lot of basic knowledge, according to the company, but it is less good than the very best models at designing experiments on its own.
Hy4 can be used through Tencent's own services, such as Tencent Cloud, WorkBuddy, CodeBuddy and the chatbot Yuanbao, and through other platforms such as OpenRouter and OpenCode. The price stays the same as for the preview: about 0.90 dollars for a million tokens (pieces of text) that you put in and 2.70 dollars for a million tokens that the model returns. A smaller, cheaper version, Hy4 Flash, is coming soon.
What does this mean for the road to the singularity?
An AI model that operates a computer on its own can do more than answer questions: it can open programs, edit files and finish a job from start to end. Until now, this skill has mostly been found in closed models from a few large companies. If the weights of Hy4 are released as promised, developers around the world will be able to build and improve such AI helpers themselves. Making things up less often matters here: anyone who lets an AI work on its own must be able to trust it. The next step is for independent tests to show how well Hy4 performs in practice.
What is the technological singularity? Read the explanation →
