Tencent Releases Hy4 With Vision and Computer Use, Open Weights Promised Later

Tencent on November 4, 2026 released Hy4, the full version of its flagship Hy language model, through Tencent Cloud, OpenRouter and its own apps. It is the first Hy model that can read images, and the first that can operate a computer, the company said.

Tencent published Hy4 preview on August 28, 2026, a mixture of experts model (one that routes each token through a small share of its weights) with 770 billion parameters in total and 49 billion active. The preview weights went out under the Apache 2.0 license on Hugging Face. The full release joins a run of Chinese model releases less than two weeks after Moonshot AI shipped Kimi K3.1 Pro.

What did Tencent release on November 4?

Tencent released Hy4 on November 4, 2026 with larger pre-training and post-training runs than the preview, image understanding and computer use, according to its announcement. The company says it also follows instructions and handles long documents and long tasks better than the preview.

Tencent says the vision features handle fine image detail and complex charts, and that for computer use it trained the model on professional software such as Blender and CAD tools. Vercel's changelog lists reasoning, tool calling, structured outputs, image input and a context window of 1 million tokens.

In its own tests, Tencent says the hallucination rate fell from 15.76% for the preview to 5.45%. In a finance test with 213 experts and 579 tasks, it says Hy4 averaged 3.04 out of 4, ahead of DeepSeek V4.1 Flash 1017 at 2.97 and Kimi K3.1 Pro and GLM 5.3 at 2.68 each.

How does Hy4 compare with Kimi, DeepSeek, GPT and Claude?

In the score table Tencent published on November 4, 2026, Hy4 has the top score on 1 of the 14 rows that are not Tencent internal tests: APEX-Agents v1.1, a test of agent work. Tencent itself writes that Hy4 still has a large gap to the top closed models, especially on its hardest internal tests in frontier math, long development tasks and visual question answering.

The full table has 34 rows; the other 20 are tests Tencent built itself and are left out below. Tencent marks its own test runs with an asterisk and the text only subset of Humanity's Last Exam with a dagger.

Counted over these 14 rows, Hy4 scores above Claude Opus 5.5 on 2 (DeepSWE and APEX-Agents) and above GPT-6.1 Sol on 5, with DeepSWE mixed. Against Kimi K3.1 Pro, Hy4 leads on 2 rows (FrontierCode 2.0 and APEX-Agents), is mixed on ProgramBench and trails on 11.

Where can Hy4 be used, and what does it cost?

Hy4 is available as of November 4, 2026 on Tencent Cloud and OpenRouter and in Tencent's WorkBuddy, CodeBuddy, Yuanbao and ima apps, at the same API price as the preview, Tencent says.

OpenRouter lists $0.834 per million input tokens and $2.501 per million output tokens, the preview's August prices, with a 50% discount for now. Vercel added Hy4 to its AI Gateway on November 3 under the name tencent/hy4, also at 50% off, through December 31, 2026.

Those list prices are well below Kimi K3.1 Pro, which Moonshot AI sells at $2 per million input tokens and $10 per million output tokens on OpenRouter. That works out to about 42% of Kimi's input price and 25% of its output price, before discounts, for a model that trails Kimi on most public rows of Tencent's own table.

When will the Hy4 weights be released?

Tencent says it plans to open source the Hy4 weights but has given no date. It has lined up support from the vLLM and SGLang projects and Inferact, and plans a system card with a fuller safety evaluation.

Until then, Tencent's claim that Hy4 ranks among the strongest open source models describes a model available only through APIs. The weights link on OpenRouter's page points to a Hugging Face repository that does not open publicly as of November 4, 2026.

What does Tencent say about Hy4's safety?

Tencent says it kept only defensive cybersecurity data in Hy4's training, such as proof of concept generation and vulnerability repair, and removed exploitation, reverse engineering and penetration testing material. An audit by its Xuanwu and Keen security labs found no prominent offensive risk, such as writing exploits, according to the company.

On biology, Tencent reports 30.4 out of 100 on its own Hy-BioExperimentBench, below what it calls its high risk threshold. Moonshot AI said on October 23 that it had removed cyber related training data from Kimi K3.1 Pro before its open release, so two Chinese labs preparing open weights this fall now say they cut offensive security data on purpose.

What we don't know yet

• When Tencent will publish the Hy4 weights, under which license, and whether the final model keeps the preview's 770 billion parameters.
• How Hy4 scores in independent evaluations rather than in Tencent's table.
• When Hy4 flash will launch and at what price.