Tencent Hy4: 770-billion-parameter AI model can see images, operate computers
By Anika SharmaAI & Innovations Editor, Peepals Tech
Representative image: danielkirsch / Pixabay
Tencent launched Hy4 on 4 November 2026, a 770-billion-parameter Hunyuan model with image understanding and computer control; it says open weights will follow soon.
What Happened
China's Tencent launched Hy4, the newest large language model from its Hunyuan team, on Wednesday, 4 November 2026. The company says Hy4 is the first Hunyuan model that can understand images, and that it can also operate a computer on a user's behalf through on-screen interfaces and command lines, while combining step-by-step reasoning with the use of software tools. Reports put the model at 770 billion parameters, up from under 300 billion in the earlier Hy3 series. Hy4 is available through Tencent's own products and cloud, and on the third-party API service OpenRouter; Tencent said the model's weights and system card would be released openly in the coming days. Tencent's Hong Kong-listed shares closed 1.1% higher at HK$401.6 on the day of the launch, according to a business news report.
Key Facts
- Launch date: 4 November 2026, by Tencent's Hunyuan team.
- Size: 770 billion parameters, against under 300 billion for the Hy3 series. Parameters are the adjustable values a model learns during training; more usually means a bigger, more capable but costlier model.
- New abilities, as described by Tencent: understanding images, and "computer use", meaning the model can click through screens and type commands to complete tasks.
- Availability: Tencent Cloud's TokenHub, the Yuanbao chatbot, WeCom and other Tencent apps, plus OpenRouter for outside developers.
- Openness: Tencent says the weights, the files that let anyone run the model on their own hardware, will follow in the coming days; until then Hy4 is used as a service.
- Company figures reported by Caixin: a hallucination rate of 7.2%, down from 12.95%, and a price of 6 yuan per million input tokens. These have not been independently tested.
- In a blind test by financial professionals, cited by one business outlet, Hy4 ranked first among open models, ahead of models from DeepSeek and Moonshot's Kimi.
Why It Matters
Hy4 shows how fast China's big technology companies are moving into "open-weight" AI, where a company publishes the trained model so that others can download, study and adapt it. One trade report described Tencent as following Alibaba, whose Qwen models took this route earlier. For developers outside China, including in India, open weights matter because a model can be run on local servers, fine-tuned for local languages and used without sending data to a foreign company's cloud.
The "computer use" feature is part of a wider shift from chatbots that answer questions to AI agents that act. A model that can read a screen and operate apps can, in principle, fill forms, move files or run software tests. That is useful, but it also raises security questions, because an agent that can click and type can also make mistakes or be misled by what it sees. Tencent's claims about accuracy and hallucination, a term for confident but false answers, are its own and have not yet been checked by independent benchmarks.
For India, the contrast in scale is clear. The same week, models built under the IndiaAI Mission were released, with the largest, Eka-1, at 26.3 billion parameters. Smaller models are cheaper to run, and two of the IndiaAI Mission models are reported to work offline on smartphones and laptops. Global releases such as Hy4 give them both a benchmark and, once weights are public, a possible base to build on.
Peepals calculation: Tencent's reported fall in hallucination rate, from 12.95% to 7.2%, is a drop of 5.75 percentage points, or about 44% (5.75 ÷ 12.95 ≈ 0.44). At 770 billion parameters, Hy4 is about 29 times the size of India's Eka-1 at 26.3 billion (770 ÷ 26.3 ≈ 29). Our calculation from the reported figures.
| Item | Reported figure |
|---|---|
| Hy4 parameters | 770 billion |
| Hy3 series parameters | Under 300 billion |
| Hallucination rate (company figure) | 7.2%, down from 12.95% |
| Input price (company figure) | 6 yuan per million tokens |
| Tencent share move on launch day | Up 1.1% to HK$401.6 |
Impact
Short-term: Developers can already test Hy4 through Tencent Cloud and OpenRouter. Tencent's own apps, including Yuanbao and WeCom, get the new model's image and agent features.
Long-term: If Tencent publishes the weights as promised, Hy4 joins a growing list of large open models from China that developers worldwide, including Indian start-ups, can adapt. That increases competition for closed models from US companies and puts pressure on prices.
Who is affected: AI developers and start-ups choosing which model to build on; companies weighing AI agents for office tasks; and Indian teams building sovereign models, which now face a larger open competitor.
What to watch next: The release of Hy4's weights and system card, which Tencent says is due in the coming days, and independent benchmark results.
Key Takeaway
Tencent's 770-billion-parameter Hy4, launched on 4 November 2026, adds image understanding and computer control to the Hunyuan family, and its promised open weights would make it one of the largest models anyone can download.
Questions and Answers
What is Tencent Hy4?
Hy4 is the latest large language model from Tencent's Hunyuan team, launched on 4 November 2026. It has 770 billion parameters, can understand images and can operate a computer through on-screen interfaces and command lines.
Is Tencent Hy4 open source?
Not yet in full. Tencent says it will release the model's weights and system card in the coming days. For now, Hy4 is available as a service through Tencent Cloud, Tencent apps and OpenRouter.
Can developers in India use Tencent Hy4?
Yes, through the OpenRouter API service, where it is listed. Once Tencent releases the weights, developers could also run and adapt it on their own servers.
How does Hy4 compare with India's own AI models?
It is far larger. Hy4 has 770 billion parameters, about 29 times the 26.3 billion of Eka-1, the largest of the models released under the IndiaAI Mission the same week.
Disclaimer: Prepared by the Peepals newsroom from publicly available sources with AI assistance. Information is accurate to the best of our knowledge at the time of publication and may change. Images may be representative. Not professional advice. Report an error via our contact page.
By Anika Sharma, AI & Innovations EditorCovers AI & Innovation and more for Peepals Tech. All stories by Anika Sharma
NICE layoffs: 500 jobs cut as AI revenue jumps 46% to $392 million
Reliance Jio Rs 10,000 plan: three years of unlimited 5G, 2.5 GB a day
Get the day's stories in your inbox
Peepals Weekly: the week decoded, free. One email a week, unsubscribe in one click.
Anika Sharma is a pen name of the Peepals Global Editorial Team, which researches, fact-checks and publishes every story. The portrait is an illustration.
