Tencent Hy4 Officially Released: Open-Weight Multimodal Flagship with 5.45% Hallucination Rate, Computer Use and 1M Context Support
Introduction
On November 4, 2026, Tencent officially released its latest flagship large language model, Hy4 — the full release following the "Hy4 Preview" version launched in August 2026. Hy4 not only significantly surpasses its predecessor on performance benchmarks but also introduces vision perception and computer use capabilities for the first time, marking an important step for Tencent in the global open-source AI race.
Against the backdrop of Chinese tech giants doubling down on AI, Hy4's release carries significant strategic importance — it is not only Tencent's flagship demonstration of AI capabilities but also a critical foundation for deep AI integration across its core business areas including finance, game development, and scientific research.
Core Technical Breakthroughs of Hy4
Large-Scale Pre-Training and Post-Training Scaling
A key characteristic of Hy4 is large-scale expansion in both pre-training and post-training phases. This dual-scaling strategy significantly improves performance on complex tasks while maintaining efficient inference.
1 Million Token Context Window
Hy4 supports a context window of up to 1 million tokens, enabling it to process:
- Entire books or lengthy technical documents
- Complete analysis of large codebases
- Meeting transcripts spanning several hours
- Complex multi-turn research conversations
First-Ever Vision Perception Capabilities
Hy4 is the first model in the Hy series to support visual perception, capable of:
- Interpreting complex financial charts and data visualizations
- Analyzing fine-grained image details
- Understanding screenshots and interface elements
- Processing complex documents mixing text and images
Computer Use Capabilities
Hy4 introduces computer use functionality, supporting:
- Graphical User Interface (GUI) control: Understanding and operating windowed applications
- Command-Line Interface (CLI) control: Executing terminal commands and scripts
- Cross-platform task automation: Coordinating complex workflows across different applications
Performance Benchmarks and Key Metrics
Dramatic Reduction in Hallucination Rate
Hy4 has achieved significant improvements in reliability:
| Metric | Hy4 Preview | Hy4 Release | Improvement |
|---|---|---|---|
| Hallucination Rate | 15.76% | 5.45% | -65.4% |
| Speculative Error Rate | 23.1% | 5.6% | -75.8% |
The hallucination rate dropped from 15.76% to 5.45%, a reduction of 65.4% — particularly significant for financial and medical applications requiring high accuracy.
Outstanding Financial Domain Performance
In evaluations involving finance professionals, Hy4 ranked first among open-weight models, surpassing competitors such as DeepSeek-V4.1-Flash-1017 and Kimi K3.1 Pro. On the internal Investment Research Agent Benchmark (iRaB), Hy4 improved by 12.03 points over the Preview version.
Game Development Capabilities
In game development benchmarks, Hy4 provides strong support for major game engines (Unreal Engine, Unity, Godot), achieving significant improvements over the Preview version in game logic and development benchmarks.
Hyra 2.0: A New Engine for Scientific Discovery
Hy4 is equipped with the upgraded Hyra 2.0 scientific research framework, supporting recursive self-improvement and scientific discovery. Released open-source alongside Hy4 are two foundational science models:
HyCell: Single-Cell Representation Model
Designed for biomedical research, HyCell performs high-precision representation learning on single-cell data, supporting cutting-edge research in genomics, proteomics, and related fields.
HyPDE: Physical Field Modeling Model
Focused on solving partial differential equations (PDEs), HyPDE is suited for physical field modeling tasks in fluid dynamics, climate simulation, materials science, and more.
The open-source release of these two models demonstrates Tencent's strategic intent to position Hy4 as scientific research infrastructure.
Availability and Pricing
Platform Integration
Hy4 has been integrated into multiple core Tencent products:
- WorkBuddy: Enterprise collaboration platform
- CodeBuddy: Code development assistant
- Yuanbao: Consumer AI assistant
- WeCom AI Assistant: Enterprise WeChat AI features
Third-Party API Access
Hy4 is also available via API through multiple third-party platforms:
- OpenRouter
- Hermes
- OpenCode
- Vercel AI Gateway
Pricing Strategy
Tencent maintains competitive pricing:
- Input tokens: ~US$0.90 per million tokens
- Output tokens: ~US$2.70 per million tokens
- Cached input tokens: ~US$0.045 per million tokens
Additionally, Tencent Cloud TokenHub and AI Gateway offer 50% discount promotions through December 31, 2026.
Positioning in the Global Open-Source AI Race
Hy4's release comes at a time when the global open-source AI race is intensifying. In the same period, several heavyweight competitors have also emerged:
- Kimi K3.1 Pro (Moonshot AI): 2.8T parameter multimodal model with 1M token context
- LongCat-2.5 (Meituan): 1.6T parameter MoE architecture focused on long-horizon agentic tasks
- Endeavor 1.1 (Flower Labs): Emphasizing auditability and data ownership
In this race, Hy4's differentiated advantages include:
- Deep integration with the Tencent ecosystem: Direct reach to hundreds of millions of users
- Professional optimization for finance: Tailored to the specific needs of Asia-Pacific financial markets
- Scientific research infrastructure: Open-sourcing HyCell and HyPDE to attract the academic community
Impact on the Asia-Pacific Region
For enterprises and developers in the Asia-Pacific region, Hy4's release carries special significance:
Language Capabilities: Hy4 has a natural advantage in Chinese language understanding and generation, making it suitable for Asia-Pacific enterprises that need to process Chinese content.
Financial Applications: In Asia-Pacific financial centers such as Hong Kong, Singapore, and Shanghai, Hy4's financial domain optimization can be directly applied to quantitative strategies, asset allocation, and investment research.
Gaming Industry: Asia-Pacific is the world's largest gaming market, and Hy4's support for Unreal, Unity, and Godot provides game developers with powerful AI assistance tools.
Scientific Collaboration: The open-source release of HyCell and HyPDE provides new computational tools for biomedical and physical science research across the Asia-Pacific region.
Conclusion
The official release of Tencent Hy4 represents another important appearance of Chinese AI technology on the global stage. By dramatically reducing hallucination rates, introducing vision perception and computer use capabilities, and open-sourcing scientific research foundation models, Hy4 demonstrates Tencent's deep technical accumulation and continuous innovation in AI.
For enterprises and developers in the Asia-Pacific region, Hy4 offers a powerful, reasonably priced, and ecosystem-rich open-source AI option. As the AI race continues to intensify, we can expect Tencent's continued investment in the Hy series and more feature optimizations tailored to Asia-Pacific market needs.
