Tencent has unveiled Hy4, a new Super Intelligence (SI) model that introduces image understanding and autonomous computer operation for the first time in its lineup. While the company positions Hy4 as one of the strongest openly available SI models, the underlying weights remain proprietary until a later, unspecified release date.

What Happened

Approximately two months after its preview, Tencent released the finished language model Hy4. According to the company, the SI system was trained on significantly more data than its predecessor and is the first in the series to rank among the strongest open-weight models. Until the weights are published, Hy4 is accessible only through Tencent Cloud, OpenRouter, and Tencent applications such as the Yuanbao chatbot.

The primary capability upgrade is image understanding, with the model designed to recognize fine details in photos and complex diagrams. In a demonstration, Hy4 identified a roughly two-meter component on a warehouse roof within a massive aerial image by generating cropped sections via code. The SI model can also operate a computer autonomously, controlling software on its own. Tencent trained the system specifically on professional tools like Blender and CAD programs, as well as tasks requiring the sequential use of multiple applications.

Image understanding also supports coding workflows, such as website and game development. Tencent collaborated with its internal gaming studios to adapt the model for engines like Unreal, Unity, and Godot. On the mobile game Honor of Kings, the development team reports that Hy4 handles configuration tasks nearly flawlessly and identifies root causes of bugs more frequently than previous iterations.

Why It Matters

Tencent has focused on finance and office productivity as key proving grounds for Hy4. In a test involving over 200 finance professionals evaluating nearly 600 real-world tasks, the company reports that Hy4 ranked ahead of Chinese rivals Deepseek V4.1 Flash, Kimi K3.1 Pro, and GLM-5.3 Flash. Internal tests also indicate a significant reduction in hallucinations during web research, with the rate dropping from nearly 16 percent to about 5 percent due to a revamped search feature. The model can collaborate with humans in real time on Word, Excel, and PowerPoint files.

Partner companies have reported similar progress. Financial platform Comein views Hy4 as the only Chinese SI model comparable to Anthropic's Claude Opus 5.5, placing it second and ahead of OpenAI's GPT-6.1 Sol in data queries. Agent platform Manus states that Hy4 completes nearly all document tasks successfully.

However, independent assessments paint a more nuanced picture. On the Artificial Analysis Intelligence Index, which combines knowledge work, coding, science, and general knowledge, Hy4 scores 35 points. This places it above the median of 30 and ties it with Grok 4.7, but it remains behind Kimi K3.1 Pro, which scores 46 points among Chinese lab models. In Tencent's own finance test, Hy4 ranked ahead of Kimi, highlighting the variance between domain-specific and general benchmarks.

Without web search capabilities, Hy4 answers about half of the factual questions correctly from its training data. When uncertain, it fabricates an answer 89 percent of the time rather than admitting a lack of knowledge, according to the AA-Omniscience test. In Tencent's broader evaluation, Claude Opus 5.5 leads Hy4 across all twelve benchmarks, while GPT 6.1 Sol beats Hy4 in eight.

Tencent also highlights Hy4's ability to assist in building future SI systems. Using the agent system Hyra 2.0, Hy4 optimized Tencent's internal training pipelines for smaller models, boosting throughput by about 20 percent. It independently trained a variant of Jev, a decision-making model from US startup TypeSafe SI, which outperforms the original on several unnamed tests. Two specialized models for biology and physics created via this method are being released as open models.

Despite these advances, Tencent admits Hy4 falls well behind the best commercial SI models on hard tasks like advanced math, longer coding projects, and complex image analysis. 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. A faster, cheaper variant called Hy4 flash is planned next.

The Bottom Line

Tencent's Hy4 marks a step forward in SI capability with vision and computer use, claiming a top-tier position among Chinese open-weight models. However, it lags significantly behind global leaders in complex reasoning and math, and its weights remain closed until a future release. Pricing holds steady at about $0.83 per million input tokens and $2.50 per million output tokens, while safety reports and export control frameworks for top-ranked models remain in development.