Tencent Launches Hy4: Vision and PC Control, Weight to Come

The new Hy4 model from Tencent arrives with image understanding and autonomous software execution. Performance varies according to benchmarks, access is available via Tencent Cloud and OpenRouter, prices remain unchanged, and the release of weights is announced for later, amid increased Chinese export controls on models.

Mixed Scores and Weight Export Under Control Project

On the Hy-FrontierMath T5 benchmark, Hy4 scores 13.0 points, compared to 97.5 for GPT 6.1 Sol and 91.3 for Claude Opus 5.5. Tencent admits that the model lags behind in advanced mathematics, extensive coding projects, and the interpretation of complex images. In the AA-Omniscience test without web access, Hy4 provides a good answer to nearly half of the questions based on its training data and generates a response in 89% of situations where it does not know the solution. Meanwhile, in October, the Chinese ministries of Commerce and Science submitted a draft regulation on exports that would include the weights of any model among the top ten globally or exceeding two trillion parameters, with an expedited procedure and automatic approval for open source. The announcement of the release of Hy4's weights is set against this backdrop.

Access via Tencent Cloud and OpenRouter, Prices Maintained

Hy4 is accessible through Tencent Cloud, OpenRouter, and Tencent applications like the Yuanbao chatbot. Tencent maintains the pricing at the preview level: approximately $0.83 per million input tokens and about $2.50 per million output tokens. A faster and cheaper variant, Hy4 flash, has been announced. The weights are expected to be released later, with no specific date provided.

Vision and Software Control Announced by Tencent

Tencent launches Hy4, presenting it as a model capable of understanding images and operating software autonomously. The company highlights its ability to recognize fine details in photos and complex diagrams, as well as to control a computer. The training included tools like Blender and CAD programs, and scenarios involving multiple applications in sequence. During a demonstration, Hy4 explored a 2.8-gigapixel aerial image by generating crops via code and identified a component approximately two meters long on a roof. Tencent states that it released the final version of the model about two months after a preview, claiming training on more data than its predecessor and positioning it among the most powerful open-weight models, while describing it as one of the best publicly available models.

Finance and Office: Internal Results and Partner Feedback

Tencent prioritizes finance and office applications. Over 200 financial experts tested Hy4 on about 600 concrete tasks; in this context, it reportedly outperformed Deepseek V4.1 Flash, Kimi K3.1 Pro, and GLM-5.3 Flash. The company also reports a decrease in the hallucination rate during online searches, dropping from about 16% to nearly 5% according to its internal evaluations, as well as the model's ability to work in real-time on Word, Excel, and PowerPoint documents. Partners observe comparable advancements: Comein ranks Hy4 on par with Claude Opus 5.5 in second place and ahead of GPT-6.1 Sol for data requests, while Manus believes it accomplishes nearly all document-related tasks. According to Tencent, Claude Opus 5.5 surpasses Hy4 in twelve aspects and GPT 6.1 Sol in eight, while Tencent's financial test ranks Hy4 ahead of Kimi. Tencent also mentions that Hy4 leads among public models in an internal research test covering seven sectors.

Independent Measures: 35 Points on the Intelligence Index

Artificial Analysis assigns 35 points to Hy4 on its Intelligence Index, above a median of 30, putting it on par with Grok 4.7. Hy4 ranks second among Chinese lab models behind Kimi K3.1 Pro with 46 points. Other independent measures place it second in China but behind leading models like Claude Opus 5.5 on mathematics and coding tasks. Without web search in the AA-Omniscience test, Hy4 answers correctly to about half of the questions and generates a response in 89% of cases when it does not know the correct answer.

Roadmap: Hyra 2.0, Jev Derivatives, and Security Before Release

Tencent indicates that it has already mobilized Hy4 with the Hyra 2.0 agent system to optimize training pipelines for smaller models, achieving a throughput gain of about 20% across six stages. The model has also trained a clean version of Jev, derived from TypeSafe AI, which Tencent claims is superior to the original in several tests, and two specialized models in biology and physics are planned for open source. Before releasing the weights, Tencent emphasizes security: filtering content related to hacking during training, no increased risk of misuse according to its labs, and a level below its internal biological risk threshold, with a detailed report announced. Previously, Tencent had released Hy3 with 295 billion parameters under Apache 2.0 in July, while competitors have advanced with GLM-5.3 (Zhipu AI), V4.1-Flash-1017 (Deepseek), and Kimi K3.1 Pro (Moonshot AI). Artificial Analysis revised its index at the end of October, now based 60% on private datasets, and measures an expanded gap between open models and top models, from 12 points compared to 3 previously.