Tencent Says Its New Hy4 AI Model Cut Hallucinations Nearly in Half

Tencent on Wednesday, Nov. 4, 2026, released Tencent Hy4, the newest version of its flagship large language model, and said it now ranks among the world’s leading open-source AI systems. The Shenzhen-based company said the model cut its hallucination rate to 7.2% from 12.95% in the earlier Hy4 Preview, adds image understanding and computer control for the first time, and will publish its model weights in the coming days.

Key figures (Tencent-reported): Hallucination rate 7.2%, down from 12.95% · $0.90 per million input tokens, $2.70 per million output · 50% off on Tencent Cloud through Dec. 31 · Weights promised “in the coming days”

What Tencent launched

According to Tencent’s announcement, Hy4 can now read images, analyze complex charts and work through visual math problems. It can also operate computers through both command-line tools and graphical interfaces, and Tencent said it trained the model on professional software including the 3D program Blender and CAD tools to handle long, multi-step jobs.

The model is live in Tencent products including its Tencent Cloud TokenHub, the WorkBuddy and CodeBuddy work assistants, the Yuanbao chatbot and the WeCom business messaging app, and its API is being offered through third-party platforms such as OpenRouter. Tencent said it plans to bring Hy4 to QQ, the game Honor of Kings, WeRead and Tencent News, and that a cheaper, faster Hy4 Flash version is coming soon.

The “open-source” label comes with a caveat: as of launch day, the final Hy4 weights were not yet posted. Tencent said it would release the weights and system cards “in the coming days.” The earlier Hy4 Preview, released in late August, is on Hugging Face under the Apache 2.0 license and lists 770 billion total parameters, 49 billion of them active for each token, with a context window of 1 million tokens. Tencent has not published those specifications for the final model.

By the numbers: what Tencent’s tests show

All of the performance figures below come from Tencent’s own evaluations. They have not been independently verified.

• Fewer made-up answers: The drop from 12.95% to 7.2% is 5.75 percentage points, a 44% relative cut. Errors caused by unsupported assumptions fell from 23.1% to 5.6%, a 76% cut (TBR calculations from Tencent figures).
• Finance test: In a blind evaluation run with Tencent Financial Technology, 213 financial professionals, 153 of them from outside institutions, scored answers on 579 real-world tasks. Hy4 averaged 3.04, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. Hy4’s lead over DeepSeek is 0.07 points, or about 2.4% (TBR calculation).
• Office work: On WorkBuddy benchmarks, Tencent reported an average score of 83.18 and a 96.15% pass rate, and said Hy4 ranked first in nine of 12 scenarios in mainland China, or 75% (TBR calculation).
• Games: Tencent said Hy4 improved 19.1% on a game-development benchmark and 9.8% on a game-logic benchmark compared with Hy4 Preview.

Price and the China open-model race

Tencent kept Hy4’s price the same as the preview’s: about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens. With the 50% Tencent Cloud discount that runs through Dec. 31, those rates fall to about $0.45 and $1.35 (TBR calculation). WorkBuddy and CodeBuddy are free from Nov. 4 to 15.

The South China Morning Post framed the launch as Tencent’s push to catch rivals DeepSeek, Z.ai and Alibaba in open-weight models. China’s open AI field has been crowded this fall, from DeepSeek’s open-source coding agent to Alibaba’s Qwen models. Tencent itself acknowledged, according to the Post, that Hy4 still trails the world’s top closed models, particularly on its hardest internal tests for research-level math, long-horizon coding, model self-improvement and advanced visual questions.

The closed-model leaders Tencent is chasing include U.S. companies such as Anthropic, whose IPO filing laid out its revenue, losses and computing costs this week.

Safety claims

Tencent said Hy4 was trained on defensive security data and that material on exploiting vulnerabilities was excluded. Experts from its Xuanwu Lab and Keen Security Lab found “no significant concerns” about the model’s offensive cyber abilities in the areas they evaluated, the company said. On biology, Tencent said the model has strong basic knowledge but is more limited than frontier models at open-ended experiment design. Outside researchers will be able to test those claims once the weights and system card are published.

Sources

• Tencent via PR Newswire: Tencent Launches Hy4, Joining the Ranks of the World’s Leading Open-Source AI Models (Nov. 4, 2026)
• South China Morning Post: Tencent unveils Hy4 as open-weight AI race with DeepSeek, Z.ai, Alibaba heats up (Nov. 4, 2026)
• Hugging Face: tencent/Hy4-preview model card
• SiliconANGLE: Tencent releases Hy4, introducing visual understanding and computer use (Nov. 4, 2026)