Images, software and game development
Tencent introduced the full version of Hy4 about two months after its preview. The company says it trained the model on a much larger volume of data than its predecessor. It has not announced a date for publishing the weights.
Hy4 can recognize fine details in photographs and complex diagrams. In one demo, the model found a component about two meters across on the roof of a warehouse shown in a large aerial image.
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
Source: the-decoder.com
The model can also operate software independently. Tencent trained it to work with professional tools, including Blender and computer-aided design systems, as well as on tasks that require using several applications in sequence.
The company also links image analysis to programming, including website and game development. For game development, Tencent worked with its own game studios and adapted Hy4 for Unreal, Unity and Godot. In Honor of Kings, the model handles parameter configuration with almost no errors and, according to the development team, finds the causes of errors much more often than its predecessor.
Finance and office tasks
Tencent also fine-tuned Hy4 for financial and office work. In a trial, more than 200 finance professionals rated the model on nearly 600 real-world tasks. Tencent says Hy4 outperformed the Chinese models Deepseek V4.1 Flash, Kimi K3.1 Pro and GLM-5.3 Flash.
The company also says that after an update to its search function, the model was less likely to make up information when working with web sources: in internal tests, the share of such answers fell from nearly 16% to about 5%. Hy4 can work with a person in real time on Word documents, Excel spreadsheets and PowerPoint presentations.
Tencent’s partners report similar results, but base their assessments on their own trials.
An independent assessment by Artificial Analysis paints a more restrained picture. On the Intelligence Index, which combines assessments of knowledge, coding, science and general knowledge, Hy4 scored 35 — above the median score of 30. That puts it level with Grok 4.7 and in second place among models from Chinese labs. Kimi K3.1 Pro ranked higher, with 46 points. In Tencent’s own financial trial, by contrast, Hy4 outperformed Kimi.
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
Source: the-decoder.com
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
Source: the-decoder.com
Without web search, results are worse. On the AA-Omniscience factual accuracy test, Hy4 answers about half of questions correctly using only its training data. When it does not know an answer, it makes one up 89% of the time instead of acknowledging uncertainty.
The model helps train other models
In an internal trial, Hy4 ranked first among open models on research tasks across seven fields. Together with the Hyra 2.0 AI agent system, it also helped optimize training pipelines for Tencent’s smaller models.
Hy4 then independently trained its own version of Jev, a decision-making model from US startup TypeSafe AI. Tencent says this version outperformed the original in several tests, but did not disclose their names.
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
Source: the-decoder.com
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
Source: the-decoder.com
The Hyra-Jev variant trained by Hy4 beats the base Jev model in six out of eight benchmarks. | Image: Tencent
Source: the-decoder.com
Tencent is releasing specialized biology and physics models created this way as open models. The company plans to use Hy4 in future to support and accelerate its own AI development.
Safety, pricing and limitations
Ahead of the planned release of the weights, Tencent outlined its safety measures. During training, the company deliberately filtered out material related to hacking. Following reviews by its cybersecurity teams, Tencent found no elevated risk of misuse; its assessment of risks in biology also remained below the threshold it had set. The company says it will publish a detailed safety report alongside the weights.
Pricing is unchanged from the preview: about $0.83 per million input tokens and about $2.50 per million output tokens. Tencent acknowledges that Hy4 still trails the best commercial models by a considerable margin on difficult math, extended programming projects and analysis of complex images. The company plans to release a faster, cheaper version called Hy4 flash next.
Last July, Tencent released Hy3, with 295 billion parameters, under the Apache 2.0 license. Since then, Chinese competitors have moved ahead: Zhipu AI released GLM-5.3, Deepseek released V4.1-Flash-1017, and Moonshot AI released Kimi K3.1 Pro.
In late October, Artificial Analysis updated its index: closed test sets now account for 60% of its score. According to the source, the gap between open models and the leaders has widened from 3 to 12 points.
Tencent’s plans to release the Hy4 weights come amid tighter rules. In October, China’s ministries of commerce and science published a draft export regulation covering weights for models in the global top 10 or models with more than two trillion parameters. For open releases, the draft provides for an expedited process with automatic approval.
Dataist’s take: Hy4 combines a broad set of capabilities with notable limitations. In particular, without search it often replaces uncertainty with a made-up answer, and it trails commercial leaders on difficult tasks. How much the model can strengthen Tencent’s position will become clearer after the weights and detailed safety report are published.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X