Tencent released Hy4 on November 4, the newest version of its flagship Hunyuan-family model, and declared that it places Hy "among the strongest open-source models for the first time." The announcement, published on Tencent's official Hy research page, describes substantial scale-ups in both pre-training and post-training compared with the Hy4 preview that the company open-sourced in August, which was released as a 770B-parameter mixture-of-experts model with a million-token context window.
Hy4 is available now on Tencent Cloud and OpenRouter, as well as inside Tencent products including CodeBuddy, WorkBuddy, Yuanbao, and ima. The company says it delivers major performance gains over Hy4 preview at the same API price. For more context on this story, see our ongoing AI news.
Vision and Computer Use Arrive in Hy
The headline capability changes are visual. Hy4 supports vision perception for the first time, letting the model understand fine-grained image details and interpret complex charts, and it supports computer use for the first time as well — operating across both command-line interfaces and graphical user interfaces.
Tencent says it trained the model on specialized software such as Blender and CAD tools, and on long-horizon workflows spanning multiple applications, to improve its ability to plan and carry out professional tasks. In developer discussions following the launch, users highlighted demonstration videos showing the model locating a small feature in a large satellite image and inferring where a photograph was taken.
Benchmarks Aimed at Real Work
Tencent's evaluation strategy leans heavily on professional domains rather than classic academic tests. In finance, the company built thousands of realistic workspaces and tens of thousands of tasks spanning fund research, equity analysis, quantitative strategies, asset allocation, corporate finance, and tax accounting. Working with Tencent Financial Technology, it had 213 finance professionals evaluate the model, 153 of them from leading financial institutions, more than 70 percent of whom reported over five years of industry experience.
Across 579 real-world financial tasks, Tencent reports that Hy4 ranks first among open-weight models with a score of 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97, Kimi K3.1 Pro at 2.68, and GLM-5.3 at 2.68. On IRAB, an internal benchmark from financial AI agent company RabyteTech, Tencent says Hy4 scores 12.03 percentage points higher than Hy4 preview.
The company also reports reliability gains: Hy4 cut the hallucination rate to 5.45 percent from 15.76 percent on Hy4 preview, and reduced speculative errors from 23.1 percent to 5.6 percent, after targeted upgrades to search covering time-sensitive retrieval, multi-source synthesis, and factual consistency verification. In scientific research, Tencent says Hy4 ranks first among open-weight models on TenureTrail, its internal benchmark of 123 research tasks across seven disciplines.
Games, Documents, and Recursive Self-Improvement
Game development received its own treatment. Tencent partnered with professional developers at its Interactive Entertainment Group — the division behind some of the world's largest game franchises — to curate training data and evaluations covering feature implementation, gameplay logic, state tracking, and asset pipelines. Tencent says Hy4 was extensively adapted for Unreal Engine, Unity, and Godot, and improves on Hy4 preview by 19.1 percent on GameDev and 9.8 percent on GameLogic, with new support for multiplayer game development.
In office workflows, Tencent improved the model's ability to co-edit Word, Excel, and PowerPoint files in real time, reporting clear blind-evaluation gains in document formatting, reliability, and usability. Manus, the AI agent platform, said in feedback included in the announcement that Hy4 achieved a task completion rate of nearly 97 percent in document creation and editing tasks.
The most speculative claim concerns recursive self-improvement. Tencent says that paired with its Hyra 2.0 harness, Hy4 optimized the company's internal pre-training infrastructure for smaller models with the same architecture and autonomously trained a variant that outperforms the official Jev baseline on multiple benchmarks. Two foundation models co-developed by Hy4 and Hyra 2.0 — HyCell for single-cell representation and HyPDE for physical field modeling — are being open-sourced to make those results reusable.
Cyber and Bio Safety Filtering
Tencent addressed capability safety directly. The company said it filtered out offensive content "such as exploit development, reverse engineering, and penetration testing to avoid amplifying potential offensive capabilities" — an approach Reuters characterized as training the model with limited cybersecurity data amid hacking fears.
On biosecurity, Tencent reports that Hy4's open-ended experimental design capabilities score 30.4 out of 100 on its Hy-BioExperimentBench, which the company places well below its high-risk threshold. Internal experts who audited the model's reasoning traces concluded that while Hy4 has a solid grasp of foundational biology, its ability to turn that knowledge into actionable experimental protocols remains limited, and Tencent says it found no evidence that Hy4 meaningfully lowers the barrier to causing biological harm.
Open Weights Promised — With a System Card
One detail complicates the "open-source" framing: the full Hy4 weights are not downloadable yet. Tencent states plainly that it plans to open-source Hy4's model weights and has partnered with the vLLM and SGLang communities to provide ecosystem support. When the weights are released, the company says it will also publish a comprehensive system card detailing its safety evaluations — a document that will be scrutinized, given that open-weight releases from Chinese labs have drawn both praise and concern this year.
Tencent also acknowledged that a gap remains between Hy4 and top proprietary models, particularly on research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering. A lighter, cheaper Hy4 Flash is planned "in the near future."
The launch lands in an increasingly crowded Chinese open-weight field, with DeepSeek, Kimi, and GLM releases all competing on price and capability. Tencent's bet is that professional-grade computer use — not raw benchmark scores — is what turns an open model into an enterprise tool.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
