Tencent releases Hy4 with vision and computer use, weights to follow

About two months after the preview, Tencent has released the finished language model Hy4. According to Tencent, it was trained on significantly more data than its predecessor and "for the first time" ranks among the strongest open-weight models. The weights themselves are not out: until they arrive at a still-unspecified date, Hy4 is available only through Tencent Cloud, OpenRouter and Tencent apps such as the Yuanbao chatbot.

The headline feature is image understanding. Hy4 is supposed to pick out fine details in photos and complex diagrams; in a demo it finds a roughly two-meter component on a warehouse roof inside a huge aerial image. It can also operate a computer for the first time, controlling software on its own. Tencent trained it on professional tools such as Blender and CAD programs, and on tasks that chain several applications together. Image understanding is also meant to help with coding, such as building websites and games. For game development Tencent worked with its own studios and adapted the model for the Unreal, Unity and Godot engines. On the mobile game Honor of Kings, the dev team says Hy4 handles configuration tasks nearly flawlessly and finds the root causes of bugs far more often than before.

Tencent put particular effort into finance and office work. In a test with over 200 finance professionals who evaluated Hy4 on nearly 600 real-world tasks, the model ranked ahead of Deepseek V4.1 Flash, Kimi K3.1 Pro and GLM-5.3 Flash, according to Tencent. Thanks to a revamped search feature, the hallucination rate during web research fell 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, but on their own internal tests: the financial platform Comein sees Hy4 as the only Chinese model on par with Claude Opus 5.5 in second place, and clearly ahead of GPT-6.1 Sol on data queries, while the agent platform Manus says Hy4 completes nearly all document tasks successfully.

An initial independent assessment from Artificial Analysis is less rosy. 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 second among Chinese lab models, behind Kimi K3.1 Pro at 46 points. Without web search the picture changes: on the AA-Omniscience factual accuracy test, Hy4 answers about half the questions correctly from its training data, and when it does not know an answer it makes one up 89 percent of the time instead of admitting the gap.

Tencent also describes Hy4 helping to build AI. 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, it optimized Tencent's own training pipelines for smaller models. It then independently trained its own version of Jev, a decision-making model from the US startup TypeSafe AI; that version outperforms the original on several unnamed tests. Tencent is releasing two specialized models for biology and physics created this way as open models. Hy4 is meant to support and speed up further AI development inside the company.

With the planned weight release in mind, Tencent stresses safety. It says it deliberately filtered hacking-related material out of training, and its 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 weights.

Pricing is unchanged from the preview: about $0.83 per million input tokens and about $2.50 per million output tokens. Tencent admits 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.

For context, Tencent released Hy3 with 295 billion parameters under the Apache 2.0 license 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, and the gap between open models and the top has widened to 12 points from 3. The planned weight release also comes as China's commerce and science ministries have published a draft export control framework in October that would cover weights of any model ranked in the global top 10 or exceeding two trillion parameters. For open-source releases the draft includes a fast-track process with automatic approval.

Key facts

Tencent released the finished Hy4 about two months after the preview, adding image understanding and computer use; the weights will follow at a still-unspecified date.
Artificial Analysis' Intelligence Index gives Hy4 35 points (median 30), tying Grok 4.7 and placing it second among Chinese lab models behind Kimi K3.1 Pro at 46.
Without web search, Hy4 answers about half of AA-Omniscience questions correctly and makes up an answer 89 percent of the time when it does not know.
Tencent says hallucinations in web research fell from nearly 16 to about 5 percent in internal tests; pricing stays at about $0.83 per million input tokens and $2.50 per million output tokens.
Tencent admits Hy4 still trails the best commercial models on advanced math, longer coding projects and complex image analysis; a cheaper Hy4 flash is planned.

Why it matters

Hy4 is Tencent's answer to Chinese rivals that have pushed ahead since Hy3 (295 billion parameters, Apache 2.0) came out last July. It adds two things the earlier model lacked: image understanding and the ability to operate software on its own, trained on tools such as Blender and CAD programs. Tencent also says Hy4 helped optimize its own training pipelines and trained a version of TypeSafe AI's Jev, a sign of the model being used inside AI development itself. The catch is that the open-weight claim rests on weights that are not yet released.

Who it affects

Developers and companies who want a Chinese model for coding, game development, finance or office document work are the target; Tencent worked with its own gaming studios and adapted the model for Unreal, Unity and Godot. Users of Tencent apps such as the Yuanbao chatbot get it directly. Teams that want to self-host must wait for the weights. Comein and Manus are named as partners reporting on it.

How to use it

For now Hy4 is available only through Tencent Cloud, OpenRouter and Tencent apps like Yuanbao. Pricing is the same as for the preview: about $0.83 per million input tokens and about $2.50 per million output tokens. The weights come at a still-unspecified date, with a detailed safety report alongside. A faster, cheaper Hy4 flash is planned next. Two specialized open models for biology and physics are also being released.

How solid is it

Most of the strong claims are Tencent's own: second place ahead of Deepseek V4.1 Flash, Kimi K3.1 Pro and GLM-5.3 Flash in its finance test, first among openly available models on its research-task test, and the drop in hallucinations from nearly 16 to about 5 percent. Comein and Manus rely on their own internal tests. The one independent figure is Artificial Analysis' initial assessment: 35 points, second among Chinese lab models, behind Kimi K3.1 Pro at 46. In Tencent's finance test Hy4 still ranked ahead of Kimi. The index was revamped in late October and is now 60 percent based on private test sets.

Risks and caveats

Without web search Hy4 invents an answer 89 percent of the time when it does not know, and gets about half of factual questions right from training data. Tencent itself says it falls well behind the best commercial models on advanced math, longer coding projects and complex image analysis. On safety, Tencent says it filtered hacking-related material and found no elevated misuse risk, but the detailed report comes only with the weights. China's draft export control framework, which would cover top 10 or over two trillion parameter models with a fast-track for open-source releases, is mentioned only as context for the planned weight release.

“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.”

— The Decoder, reporting on Tencent