Tencent Hy4 AI Model Achieves 7.2% Error Rate with New Visual Understanding and Computer Use Features

Tencent Hy4 Introduces Visual Understanding and Computer Use Capabilities

Tencent has launched Hy4, a new open-source AI model that marks a significant advancement in the company's AI portfolio by introducing visual understanding and computer use capabilities for the first time1

1

2

2

. This large language model weighs in at 770 billion parameters with 49 billion active parameters and provides a 1 million token context window, positioning it alongside current frontier LLM offerings from European and Chinese developers2

2

. The model can process detailed images, interpret complex charts, solve mathematical problems involving visual information, and interact with both command-line tools and graphical interfaces1

1

. Tencent plans to release Hy4's model weights and system cards in the coming days as part of its commitment to responsible open-source AI development2

2

.

Hy4 Reduces Hallucination Rate and Improves Instruction Following

Tencent substantially scaled both pre-training and post-training for Hy4, delivering gains over the Hy4 Preview version that arrived on August 282

2

. Internal evaluations revealed a dramatic reduction in the hallucination rate, falling from 12.95% to 7.2%1

1

. Errors linked to unsupported assumptions dropped even more significantly, from 23.1% to 5.6%1

1

. The AI model demonstrated improvements in instruction following, long-context understanding, agentic long-horizon tasks, and deep research capabilities2

2

. Tencent identified and addressed issues with excessive reasoning and over-verification during the previous version's thinking processes2

2

.

Financial Analysis Performance Leads Industry Evaluations

Hy4 has been optimized for professional financial analysis through training on thousands of specialized workspaces and tens of thousands of high-quality tasks derived from real-world scenarios3

3

. The model gained deep understanding of expert workflows in fund research, securities analysis, and quantitative investment strategies3

3

. In a blind evaluation involving 213 financial professionals across 579 real-world financial tasks, Hy4 achieved the highest overall score of 3.04, narrowly ahead of DeepSeek-V4.1-Flash-1017 at 2.971

1

. Approximately 70% of participating professionals had over five years of industry experience, lending credibility to the assessment1

1

. This performance positions Hy4 as a competitive option for financial institutions seeking AI-powered analytical capabilities.

Workplace Productivity and Game Development Applications

Tencent designed Hy4 to unlock new opportunities for workplace productivity and collaboration. The model achieved an average evaluation score of 83.18 on its WorkBuddy platform with a 96.15% pass rate across evaluated tasks1

1

. It ranked first in nine of 12 tested scenarios within mainland China1

1

. WorkBuddy integrates Hy4 with Tencent's workplace collaboration software WeCom, allowing the model to access information across chats, emails, documents, calendars, and meetings1

1

. For game development, Hy4 improved its GameDev benchmark score by 19.1% over Hy4 Preview, with GameLogic scores increasing by 9.8%2

2

. The model integrates deeply with major engines including Unreal Engine, Unity, and Godot, supporting complex game mechanics and multiplayer implementations1

1

2

2

.

Cybersecurity Safeguards and Pricing Strategy

Tencent incorporated cybersecurity safeguards and biosafety measures into Hy4's design. The model was trained on defensive security data, including vulnerability analysis and remediation, but excludes offensive exploitation materials1

1

3

3

. In terms of biosafety, Hy4's capabilities in experimental design remain limited compared to other frontier models3

3

. Despite significant enhancements over Hy4 Preview, Tencent maintained competitive pricing at approximately $0.90 per million input tokens, $2.70 per million output tokens, and $0.045 per million cached input tokens1

1

2

2

. Hy4 is currently available on Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI Assistant, and third-party services including OpenRouter and OpenCode2

2

. Tencent plans to release Hy4 Flash soon, targeting applications where lower operating costs matter most1

1

. This release comes amid reports that Oracle terminated a $7 billion agreement to provide AI computing capacity to Tencent after U.S. government officials raised concerns3

3

.