Tencent’s Hy4 can interpret images and control software, expanding the company’s model into tasks that involve more than text. Tencent says it has also used Hy4 to improve training pipelines for smaller models. Its weights are expected later; for now, the model is available through Tencent Cloud, OpenRouter and Tencent apps such as Yuanbao.
From image analysis to computer tasks
Hy4 is Tencent’s first model in this line with image understanding. The company says it can identify fine details in photographs and complex diagrams. In one demonstration, it located a roughly two-meter component on a warehouse roof within a large aerial image.
The model can also operate a computer, using software without a person carrying out each step. Tencent says it trained Hy4 for professional tools including Blender and CAD programs, as well as workflows that require moving between multiple applications.
These capabilities are intended to support coding and development work, too. Tencent worked with its gaming studios to adapt Hy4 for Unreal, Unity and Godot. The team behind Honor of Kings says the model handles configuration tasks nearly flawlessly and identifies the root causes of bugs more often than before.
Finance and office work are key tests
Tencent says it gave particular attention to finance and office applications. In its evaluation, over 200 finance professionals assessed Hy4 on nearly 600 real-world tasks. The company reports that it ranked ahead of Deepseek V4.1 Flash, Kimi K3.1 Pro and GLM-5.3 Flash in that test.
A revised search feature is meant to reduce unsupported answers during web research. In Tencent’s internal tests, the reported hallucination rate fell from nearly 16 to about 5 percent. Hy4 is also designed to work with people in real time on Word, Excel and PowerPoint files.
Partner companies have shared positive results from their own evaluations. Comein says Hy4 is the only Chinese model on par with Anthropic’s Claude Opus 5.5 in second place, and ahead of OpenAI’s GPT-6.1 Sol on data queries. Manus says Hy4 completes nearly all document tasks successfully. These are partner assessments, rather than independent comparisons.
Independent results show strengths and limits
Artificial Analysis gives Hy4 35 points on its Intelligence Index, which combines knowledge work, coding, science and general knowledge. That score is above the median of 30, ties Hy4 with Grok 4.7, and puts it second among Chinese lab models, behind Kimi K3.1 Pro at 46 points. Tencent’s finance evaluation, by contrast, placed Hy4 ahead of Kimi.
Hy4’s results are weaker on a factual accuracy test without web search. On AA-Omniscience, it answered about half the questions correctly from its training data. When it did not know an answer, it made one up 89 percent of the time instead of acknowledging the gap. The contrast suggests that its search feature matters when the model needs to ground answers in information beyond its training.
Tencent also says Hy4 ranked first among openly available models in an internal research evaluation spanning seven fields. Used with the agent system Hyra 2.0, it helped optimize the company’s training pipelines for smaller models. Tencent says it also independently trained a version of Jev, a decision-making model from TypeSafe AI, that outperformed the original on several unnamed tests. Two resulting models for biology and physics are being released as open models.
Weights, pricing and remaining gaps
Tencent has not yet announced when Hy4’s weights will be released. The company says it filtered hacking-related material from training and that its security labs found no elevated risk of misuse. It also says the model is below its risk threshold for biological hazards; a detailed safety report is planned alongside the weights.
The current price is the same as for the preview: about $0.83 per one million input tokens and about $2.50 per one million output tokens. Tencent acknowledges that Hy4 remains well behind the best commercial models on advanced math, longer coding projects and complex image analysis. A faster, cheaper version called Hy4 flash is planned.
For now, Hy4’s standing depends on the task and the test. It combines image understanding, software control and office workflows, while its independent evaluation shows room to improve on difficult reasoning and factual reliability when search is unavailable. The eventual release of its weights will make the model more broadly available, but its performance claims still need to be read alongside those limits.