Tencent unveils Hy4 open-source AI model

On November 4, Tencent Hunyuan officially released its new-generation large language model Hy4; Tencent said the model joins the top tier of Chinese open-weight AI systems and breaks into the top tier of global open-source models.

Hy4 has a total parameter scale of 770 billion and adopts a mixture-of-experts architecture. Tencent plans to open-source the model weights and notes that the new release further expands pre-training and post-training compared with the earlier Hy4 preview, with a focus on real productivity tasks.

Hy4 introduces visual understanding capabilities for the first time in the Hunyuan series, enabling it to recognize details in images, analyze complex charts, and complete visual math tasks with reasoning. The model is also designed to support scenarios such as educational web design and analysis of promotional charts.

The model additionally debuts Computer Use, allowing it to operate graphical interfaces and command lines to execute complex workflows. Tencent Hunyuan has carried out specialized training on professional software such as Blender and CAD, as well as cross-application tasks that require long-range coordination.

Hy4 has been evaluated across finance, office, development, and scientific research scenarios; Tencent said it is among the strongest open-source models.

In a crowdsourced financial test involving 213 financial experts and 579 real tasks, Hy4 scored 3.04, higher than several named comparative models. For financial use cases, Tencent has optimized timeliness of retrieval, multi-source information integration, and factual consistency verification, with internal tests indicating a reduction in hallucination rate to 7.2% and a drop in guessing-related error rate to 5.6% compared with the preview.

In front-end and game development, Hy4 shows improved capabilities in complex interactions, business logic, and state management, and enhanced compatibility with major game engines including Unreal Engine, Unity, and Godot. Internal evaluations report gains over the preview version in GameDev and GameLogic benchmarks.

Tencent said internal TenureTrail evaluations, covering seven disciplines and 123 executable research tasks, showed the model had strong task-handling ability.

Hy4’s training data for cybersecurity was restricted to defensive use cases such as proof-of-concept generation and vulnerability remediation, excluding offensive topics like exploitation and penetration testing. Tencent said internal audits by security experts found Hy4 did not exhibit significant offensive cybersecurity capabilities.

Tencent Hunyuan is also preparing for Recursive Self-Improvement, using Hy4 and the Hyra 2.0 framework to optimize pre-training infrastructure for smaller models and automate parts of model training.

Hy4 continues Tencent Hunyuan’s development pace of releasing major versions roughly every two months since 2026, scaling from the Hy3 series with fewer than 300 billion parameters to the current 770 billion. Tencent states that the next generation Hy5 will further increase in scale.

Despite performance improvements over the preview, Hy4 maintains the same pricing: 6 yuan per million tokens for input, 18 yuan per million tokens for output, and 0.3 yuan per million tokens for cache hits. Tencent emphasizes model practicality and cost-effectiveness while expanding training investment.

Hy4 is already available through Tencent Cloud Tokenhub and products such as WorkBuddy, CodeBuddy, Yuanbao, WeChat Work Smart Assistant Dayuan, and ima, with APIs integrated into platforms including OpenRouter, Hermes, and OpenCode. Tencent services such as QQ, Honor of Kings, Peace Elite, WeChat Reading, and Tencent News are in the process of integrating the model.