Tencent Hunyuan has officially released Hy4, which for the first time introduces visuals and supports computer operations compared to the preview at the end of August.
At the same time, the model's performance has significantly improved over its predecessor, such as allowing it to create a game by itself and play with its computer, which is quite astonishing. Here is the result:
The usual rule: This week at AGI Bar, Hy4 can be used for free
Context window of 1M tokens, with a maximum output of 64K at once, the API is priced the same as the preview, with every million tokens input costing 6 yuan (0.3 yuan for cache hits) and output costing 18 yuan.
According to the release evaluation from Agent Arena, Hy4 achieved a net improvement of 4.7% in over 8500 real long-range agent tasks, ranking 14th overall and entering the Pareto frontier: Hy4 has a cost of $0.55 per task, about 62% lower than Kimi K3.1 Pro (Max) at $1.46.
Swipe left or right to view · Total of 2 images
The full benchmark results are as follows:
If you want to use it directly, open WorkBuddy and switch the model to Hy4. Free quotas will be given daily from November 4 to 15, and subscription users will have higher quotas.
In addition, it can also be used through yuanbao, ima, and API calls can be made through Tencent Cloud and OpenRouter.
I will provide more examples later, all generated by Hy4 based on specific requirements, such as this piece of blue-and-white porcelain.
Better visuals, better agent
The entrance to professional software is mostly graphical interfaces. Hy4 integrates visuals, reasoning, and tool calls, and has also undergone special training for Blender, CAD, and cross-application long tasks.
For example, if you give it a casually taken photo, it can locate the city and shooting point.
This capability of Computer use can also be applied to interior design, such as the example below: modeling the interior in Blender according to a blueprint, then importing it into UE to create a first-person interactive renovation scene with material switching.
Hy4 also has excellent adaptability for tools like UE, Unity, Godot, and this example is "QQ Happy Arcade Hall" made with Three.js.
I tried it out
Besides the initial two-player tank battle, I also created more examples and will share them with everyone.
Planetary gear reducer: "Build from scratch using Blender's Python script, calculating gear shapes using involute curves, ensuring the gears mesh properly, and create an animation where it operates first, then explodes apart."
Blue-and-white porcelain throwing: "Create a web page for throwing pottery: pressing down on the clay allows it to rise and shape, glaze it after shaping, paint the intertwined lotus with blue-and-white material, and then fire it in a kiln."
Understanding the Qingming Festival River Painting: "This is the complete volume of the Qingming Festival River Painting from the Qing dynasty edition, 56531×1700 pixels. Slice it and view it section by section without internet, find the 12 most interesting details in the market scene, frame them accurately, and create a scroll guide."
Hy4 sliced it into 24 sections for viewing; I checked the 12 frames I cropped, with the camel caravan, cold food swing, and high platform performance all accurately placed.
The dimensional paper-cutting book of "The Eight Immortals Crossing the Sea" (one-sentence requirement):
Milk tea chain business analysis: "This is a sales breakdown for a milk tea chain from January to October, and the owner needs to decide where to open stores next year, whether to continue discounts for takeout, and which products to cut. Deliver an Excel model with formulas and a PPT of 20 pages or less."
The data was compiled by 26,000 lines of script I wrote, deliberately embedding two changes: the summer surge of lemon tea and the decline of Ice Bock thick milk starting in July. Hy4 discovered them and suggested testing before cutting Ice Bock.
The most training in finance
This model has also undergone specialized training in finance, covering directions such as fund research, securities analysis, quantitative strategies, asset allocation, and tax accounting.
In Hunyuan's self-built tests, the hallucination rate and error rate caused by guessing have significantly decreased.
Tencent Fintech FiT organized a survey with 213 experts, 153 of whom came from leading external financial institutions, with 70% having over five years of experience. The results of 579 real finance tasks (with a full score of 4) are:
Hy4:
3.04DeepSeek-V4.1-Flash-1017:
2.97Kimi K3.1 Pro:
2.68GLM-5.3:
2.68
This is a Tesla valuation report created with Hy4, outputting an Excel model and PDF.
Business replaces the model to set questions
Yao Shunyu and Tang Daosheng mentioned in their discussion that after the methodology matures, finding good questions becomes harder; therefore, why not take real tasks to see? Here are several internal evaluation results from the business:
Honor of Kings: The accuracy rate of 20 types of game configurations improved from 91.74% to 97.28%, and the Agent's autonomous verification correction reached
99.32%.QQ: The overall satisfaction rate of Little Q improved from 90.11% to 93.10%.
Yuanbao: Internal evaluation scores increased from 65.25 to 72.53.
WeChat Reading: Comprehensive evaluation scores rose from 82.10 to 87.69.
WorkBuddy: An average score of 83.18 with a pass rate of 96.15%, with 9 out of 12 scenarios ranked first among domestic models.
After connecting to WorkBuddy's enterprise WeChat suite, Hy4 can link chat records, emails, documents, schedules, and meetings to clarify their respective projects.
Let the model participate in training the model
Hy4 is preparing for Recursive Self-Improvement (RSI), in conjunction with the automated research scaffolding Hyra 2.0:
On a scaled-down internal model, it self-locates performance bottlenecks in pre-training systems, improving end-to-end training throughput by
20.1%.Automatically trains Hyra-Jev, outperforming the official Jev in 6 out of 8 benchmark evaluations.
In research tasks, Hy4 completed the three-dimensional reconstruction of mouse embryos with 84 unordered sections and 6.99 million cells without any specialized registration tools or cell type annotations, all without intervention, taking 1.5 hours.
In collaboration with the mathematics team from Peking University, Hy4 found a counterexample to the Méndez–Pinto–Villarreal conjecture, proposed new conditions to improve the conjecture, and conducted formal verification in Lean.
The single-cell model HyCell and the physical field model HyPDE, developed in collaboration with Hyra 2.0, have also been open-sourced.
Swipe left or right to view · Total of 2 images
Finally
The Hunyuan project plans to open-source weights and has already partnered with the vLLM, SGLang community, and Inferact for ecological support, and will release system cards together at that time.
Next, there will also be Hy4 flash, to be released soon, focusing on cost-effectiveness and reasoning speed.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。
