Tencent released Hy4, its latest large language model, on November 4, 2026. The company said it ranks among the world’s leading open-source models.
It is the first time Tencent’s models have joined the top tier of Chinese open-weight systems. Open-weight means the trained settings of the model are published so others can run it themselves. Tencent says Hy4 matches Z.ai, Moonshot AI and DeepSeek in tests ranging from coding to agents and browser use. The API costs about $0.90 per million input tokens and is available through OpenRouter, Hermes and OpenCode.
Chinese media reported that Tencent’s capital spending rose 176% year on year to 52.8 billion yuan (around $7.9 billion) in the second quarter.
However, Hy4 still trails the world’s top proprietary models. Tencent named gaps in research-level maths, long-horizon coding, model self-improvement and advanced visual question answering. In its release note the company wrote: “We will continue to steadily scale pre-training and post-training to narrow this gap.”
Who Ran the Tests That Put Hy4 Ahead?
Tencent’s Hy team and Tencent Financial Technology ran a blind finance evaluation with 213 professionals and 579 tasks, the company said. Hy4 scored 3.04. DeepSeek-V4.1-Flash-1017 scored 2.97. Both Kimi K3.1 Pro and GLM-5.3 scored 2.68.
| Model | Score |
|---|---|
| Hy4 | 3.04 |
| DeepSeek-V4.1-Flash-1017 | 2.97 |
| Kimi K3.1 Pro | 2.68 |
| GLM-5.3 | 2.68 |
Benchmarks in Tencent’s own technical blog show Hy4 ahead of Z.ai’s GLM-5.3 and DeepSeek-V4.1-Flash-1017 on all tested tasks. Against Moonshot’s Kimi K3.1 Pro the two trade wins.
Tencent’s internal evaluations put Hy4’s hallucination rate at 7.2%, down from 12.95% in Hy4 Preview. On WorkBuddy tasks Hy4 scored an 83.18 average with a 96.15% pass rate and ranked first in nine of twelve mainland China scenarios.
We reported in April that Tencent was reportedly pushing for a stake of as much as 20% in DeepSeek, which was said to push back.
What Does Hy4 Add Over Hy3 and Hy4 Preview?
Hy4 adds visual understanding for the first time, Tencent said. This covers image processing, chart analysis and mathematical visual reasoning.
It also introduces computer use for the first time. The model can operate command-line and graphical interfaces and was trained on Blender and CAD software.
The Hy4 series has 770 billion parameters. That is 2.6 times the 295 billion in Hy3. Hy4 Preview launched in late August and ranked 23rd on Code Arena’s WebDev leaderboard on November 4.
On safety, Tencent said it excluded offensive exploit material from Hy4’s training. Its own security labs found no significant offensive-cyber concerns in the areas evaluated. Internal expert reviews found strong biological knowledge but weaker open-ended experimental design than leading models.
When Can Developers Get the Hy4 Weights?
Tencent plans to release Hy4’s weights and system cards in the coming days, its release said.
The release also does not give a European availability, data-residency or EU AI Act position for Hy4. For how European models compare see European LLMs in 2026: The Complete Map and What Is the Best Open-Source LLM in 2026?.
The Hy4 API is available through Tencent products such as Yuanbao and WeCom. Output costs $2.70 per million tokens and cached input costs $0.045, the same as Hy4 Preview. Hy4 is 50% off the standard price on Tencent Cloud TokenHub through December 31, 2026.
Tencent said a lighter, cheaper Hy4 Flash is due soon and that it will continue scaling pre-training and post-training.
Author: Grace Sharp
See Also:
Chinese LLMs in 2026: The Complete Map
