Tencent releases Hy4: AI that operates a PC, but fabricates answers to 89% of unfamiliar questions without search

Tencent has released Hy4, a language model that can understand images for the first time and operate programs on its own: it can open Blender, CAD editors, Word, Excel and PowerPoint. But there is an important caveat. In the independent Artificial Analysis ranking it scored 35 points (the market median is 30).

That puts it among the strong but not the best models, and only second among Chinese ones. And when Hy4 has no web search access and doesn't know the answer, in 89% of cases it doesn't admit it. It makes something up.

Imagine asking it about a configuration parameter of a library that wasn't in its training data: with search, it will most likely find the documentation; without search, it will confidently write an option that doesn't exist. The practical takeaway is simple: try Hy4 where it has sources and tools, and don't trust its memory.

What Hy4 can do

The main new feature is image understanding. The model used to work only with text; now it parses photos and complex diagrams. In a demo, Hy4 searched a 2.8-gigapixel aerial image for an object about two meters across on a warehouse roof. To do this, it wrote code to cut the image into fragments and went through them one by one, the way a person zooms in on part of a map.

The second new feature is computer control. The model was trained on professional tools such as Blender and CAD programs, as well as on tasks that require working through several applications in sequence. It can also edit Word, Excel and PowerPoint files together with a human in real time.

For game developers, Tencent adapted the model to Unreal, Unity and Godot. According to the team, in the mobile game Honor of Kings Hy4 handles configuration tasks almost without errors and finds the root causes of bugs much more often.

What Tencent says and what independent tests show

According to Tencent itself, in a test involving more than 200 finance professionals and almost 600 real-world tasks, Hy4 outperformed the Chinese models Deepseek V4.1 Flash, Kimi K3.1 Pro and GLM-5.3 Flash. The share of fabricated facts in web research fell from almost 16% to about 5% in internal measurements, thanks to a reworked search system. Partners, including the financial platform Comein and the agent platform Manus, also praise the model, but they rely on their own tests. In Tencent's evaluation, Claude Opus 5.5 beats Hy4 on all twelve benchmarks, and GPT 6.1 Sol beats it on eight.

The independent picture is more modest. The Intelligence Index from Artificial Analysis combines knowledge work, coding, science and general knowledge. There Hy4 scores 35 against a median of 30, the same as Grok 4.7. The leader among Chinese models, Kimi K3.1 Pro, scored 46. Roughly on par with Hy4 is Anthropic's smallest model, Haiku 5.5, which costs almost half as much per benchmark task.

In the AA-Omniscience factual accuracy test, Hy4 answers correctly about half of the questions on material it memorized during training. On hard math the gap is huge: in Hy-FrontierMath T5 it scores 13.0 versus 97.5 for GPT 6.1 Sol and 91.3 for Claude Opus 5.5. Tencent itself admits that on long coding projects and complex image analysis the model falls well behind the best commercial models.

A model that helps train other models

Together with the Hyra 2.0 agent system, Hy4 optimized the internal infrastructure Tencent uses to train small models: throughput rose by about 20% across six stages. The model then independently trained its own version of Jev, a decision-making model from the US startup TypeSafe AI. This version beat the original on six of eight benchmarks. Tencent will release two specialized models for biology and physics built this way as open source. These claims cannot yet be verified externally.

Safety, price and context

Tencent promises to open the Hy4 weights, but has not given a date. For now, access is only through Tencent Cloud, OpenRouter and Tencent apps, such as the Yuanbao chatbot. The company says it filtered hacking-related material out of the training data, its labs found no elevated risk of misuse, and biological risks sit well below the threshold. A detailed safety report will come out together with the weights.

The price stays the same as in the preview: about $0.83 per million input tokens and $2.50 per million output tokens. A faster and cheaper version, Hy4 flash, is planned next.

Hy3, with 295 billion parameters, came out under the Apache 2.0 license in July, but Chinese competitors have since pulled ahead: GLM-5.3, Deepseek V4.1-Flash-1017, Kimi K3.1 Pro. The gap between open models and the leaders has grown from 3 to 12 points. In October, Chinese authorities published draft export control rules for model weights from the global top 10 or with more than two trillion parameters; open releases get a fast-track procedure with automatic approval.

What to do with this

Hy4 is worth testing in office automation, financial analytics, document work and game development tools, anywhere it has search and access to sources. For hard math, long coding projects and answers "from memory" without sources, the top closed models remain more reliable. Re-check the figures from Tencent and its partners on your own data: independent evaluations are noticeably more modest.