Tencent has unveiled its AI model Hy4. It can understand images for the first time and operate software on its own. The company counts it among the best openly available models, though the model weights won't be released until later.
About two months after the preview, Tencent has released the finished language model Hy4. According to Tencent, the model was trained on significantly more data than its predecessor and "for the first time" ranks among the strongest open-weight models. Until the weights are released at a still-unspecified date, Hy4 is only available through Tencent Cloud, OpenRouter, and Tencent apps like the Yuanbao chatbot.
The biggest new feature is image understanding. Hy4 is supposed to recognize fine details in photos and complex diagrams. In a demo, it finds a roughly two-meter component on a warehouse roof inside a massive aerial image. The model can also operate a computer for the first time, controlling software on its own. Tencent trained it specifically on professional tools like Blender and CAD programs, as well as tasks that require using multiple applications in sequence.
Hy4 searches a 2.8-gigapixel aerial image for a roughly two-meter component on a rooftop by generating cropped sections via code. | Image: Tencent
Image understanding is also meant to help with coding, such as building websites and games. For game development, Tencent worked with its own gaming studios and adapted the model for popular game engines like Unreal, Unity, and Godot. On the mobile game Honor of Kings, Hy4 handles configuration tasks nearly flawlessly and finds the root causes of bugs far more often than before, according to the dev team.
Finance as a key proving ground
Tencent put particular effort into finance and office applications. In a test with over 200 finance professionals who evaluated Hy4 on nearly 600 real-world tasks, the model ranked ahead of Chinese rivals Deepseek V4.1 Flash, Kimi K3.1 Pro, and GLM-5.3 Flash, according to Tencent. Hy4 is also supposed to hallucinate far less during web research. Thanks to a revamped search feature, the hallucination rate dropped from nearly 16 to about 5 percent in internal tests. The model can also collaborate with humans in real time on Word, Excel, and PowerPoint files.
Partner companies report similar progress, though they're also relying on their own internal tests. Financial platform Comein sees Hy4 as the only Chinese model on par with Anthropic's Claude Opus 5.5 in second place, and clearly ahead of OpenAI's GPT-6.1 Sol on data queries. Agent platform Manus reports that Hy4 completes nearly all document tasks successfully.
In Tencent's evaluation, Claude Opus 5.5 leads Hy4 across all twelve benchmarks. GPT 6.1 Sol beats Hy4 in eight of them. | Image: Tencent
An initial independent assessment from Artificial Analysis paints a slightly different picture. On its Intelligence Index, which combines knowledge work, coding, science, and general knowledge, Hy4 scores 35 points, above the median of 30. That ties it with Grok 4.7 and puts it in second place among Chinese lab models, behind Kimi K3.1 Pro at 46 points. In Tencent's own finance test, Hy4 still ranked ahead of Kimi.
Without web search, though, the picture changes. On the AA-Omniscience factual accuracy test, Hy4 answers about half the questions correctly from its training data. When it doesn't know an answer, it makes one up 89 percent of the time instead of admitting the gap.
On the Artificial Analysis Intelligence Index, Tencent's model trails the strongest Chinese model, Kimi 3.1 Pro, by a wide margin. It roughly ties with Anthropic's smallest model, Haiku 5.5, which costs almost half as much per task on the benchmark. | Image: Artificial Analysis
AI that helps build the next version of itself
In an internal test with research tasks across seven fields, Hy4 ranks first among openly available models, according to Tencent. Together with the agent system Hyra 2.0, Hy4 also optimized Tencent's own training pipelines for smaller models.
According to Tencent, Hy4 with the Hyra 2.0 harness boosts throughput of the company's internal training infrastructure by about 20 percent across six steps. | Image: Tencent
It then independently trained its own version of Jev, a decision-making model from US startup TypeSafe AI. That version outperforms the original on several unnamed tests. Tencent is releasing two specialized models for biology and physics that were created this way as open models.
The Hyra-Jev variant trained by Hy4 beats the base Jev model in six out of eight benchmarks. | Image: Tencent
Going forward, Hy4 is meant to support and speed up further AI development inside the company.
Safety, pricing, and admitted weaknesses
With the planned weight release in mind, Tencent is stressing safety. The company says it deliberately filtered out hacking-related material during training, and Tencent's own security labs found no elevated risk of misuse. The model also falls well below Tencent's risk threshold for biological hazards. A detailed safety report will follow alongside the model weights.
Despite the improvements, Tencent is charging the same price as for the preview. One million input tokens costs about $0.83, and one million output tokens runs about $2.50. The company admits, however, that Hy4 still falls well behind the best commercial models on hard tasks like advanced math, longer coding projects, and complex image analysis. A faster, cheaper variant called Hy4 flash is planned next.
The gap is widest on Hy-FrontierMath T5, where Hy4 scores 13.0 points compared to 97.5 for GPT 6.1 Sol and 91.3 for Claude Opus 5.5. | Image: Tencent
Tencent released Hy3 with 295 billion parameters under the Apache 2.0 license just last July. Since then, Chinese rivals have pushed ahead. Zhipu AI shipped GLM-5.3, Deepseek released V4.1-Flash-1017, and Moonshot AI launched Kimi K3.1 Pro. Artificial Analysis revamped its index in late October, with 60 percent of it now based on private test sets. The gap between open models and the top has widened to 12 points from 3.
The planned release of Hy4's weights comes at a time of tighter rules. China's commerce and science ministries published a draft export control framework in October that would cover model weights for any model ranked in the global top 10 or exceeding two trillion parameters. For open-source releases, however, the draft includes a fast-track process with automatic approval.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
原文版权归 The Decoder 及原作者所有，本站转载仅供阅读。如需删除，请联系 support@aisa.one。
Studio
