Tencent promises open weights "in the coming days" after API-only launch of Hy4

Tencent promises open weights "in the coming days" after API-only launch of Hy4

On November 4, Tencent launched its flagship model Hunyuan Hy4, a sweeping upgrade of August's Hy4 Preview with first-generation multimodal understanding and the first computer-use capabilities in the Hunyuan series.

What Hy4 actually does

Hy4 is, according to Tencent's announcement as reported by WinBuzzer, the company's first Hunyuan model with vision. More concretely, it means the model gets the first computer-use capabilities in the series: it can work across command-line environments and graphical user interfaces to perform both professional and everyday tasks.

Targeted training covers, according to Tencent's announcement as reported by WinBuzzer, specialized software such as Blender and CAD applications, together with long-horizon tasks spanning multiple applications — something Tencent highlights as an improved ability to plan and execute professional workflows. The company also cites adaptations for game development in UE, Unity and Godot, with reported improvements of 19.1% on the GameDev evaluation and 9.8% on GameLogic compared with the August preview version.

The numbers — and who stands behind them

All the result figures below are Tencent-reported and relayed via secondary sources. None have been independently reproduced, and Bloomberg News says explicitly that it has not verified the company's benchmarks.

On the computer-use benchmark OSWorld 2.1, Tencent's table gives 78.9 points for Hy4, versus 81.8 for Claude Opus 5.5, 79.6 for Kimi K3.1 Pro and 75.2 for GPT-6.1 Sol, according to Startup Fortune. Hy4 thus sits below the top in the company's own comparison — and the competitor models' results are just as unverified as its own.

For the WorkBuddy platform, Tencent reports an average evaluation score of 83.18 and a pass rate of 96.15% for Hy4, as relayed in a secondary account of the company's figures.

In finance, Tencent highlights a blind evaluation with 579 realistic financial tasks and 213 professionals. According to WinBuzzer, 153 of the experts came from leading financial institutions outside Tencent, and around seven in ten had more than five years of experience. A secondary account states that Hy4 received the highest overall score, and concrete figures are also claimed against competitor models such as DeepSeek-V4.1-Flash-1017, Kimi K3.1 Pro and GLM-5.3 — models whose versions and existence cannot be verified from the available documentation, and which are relayed here only "as named in Tencent's evaluation".

Two conflicts that cannot be resolved

The available coverage gives conflicting answers on two points. First, the hallucination rate: one secondary account states that internal evaluation showed a drop from 12.95 to 7.2 percent, while Startup Fortune states that Tencent's internal search evaluations showed a drop from 15.76 percent in the August preview to 5.45 percent in this release. The figures may stem from different internal evaluation setups, but this cannot be confirmed from the available basis — both should therefore be read as Tencent-reported and not mutually reconcilable.

Second, the price. Startup Fortune reports that after a free usage quota from November 4 to 15, the model costs $0.834 per million input tokens and $2.501 per million output tokens. Another account gives roughly $0.90 for input and $2.70 for output per million tokens. These may be rounding differences, but they cannot be confirmed.

Safety and training on defensive-only content

Reuters reports that Tencent launched Hy4 on Wednesday, November 4, with cybersecurity data limited to "defensive" content, to reduce the risk that the model could contribute to cyberattacks. Bloomberg describes the same move as similar to the training-data controls Moonshot has described for Kimi K3.1 Pro. This is a company statement about training data, not something that can be independently verified from the available documentation.

The Hunyuan series is led, according to Bloomberg, by former OpenAI researcher Yao Shunyu. Bloomberg describes how, since arriving at Tencent the summer before, he has rebuilt the research team and changed a development culture he himself had publicly criticized for placing too much weight on "playing" benchmarks. That is a remarkable backdrop against the company now presenting a series of its own headline evaluation figures that no independent party has yet reproduced.

The business context: enterprise agents as a revenue source

Tencent cloud chief Dowson Tong said, according to Bloomberg, in an interview last week that the company is betting on enterprise agents to better monetize AI. A bright spot in the portfolio is the AI office tool WorkBuddy, which Bloomberg describes as the clear leader in the field, with 32 million monthly active users in September — more than direct rivals combined, according to the news agency.

Tencent further states, as relayed by WinBuzzer, that WorkBuddy was launched in March 2026, has rolled out over 100 major version updates, increased token consumption 50-fold from the March launch period, and completed over 9 billion task requests to date. These too are company-provided figures. The connection is nonetheless clear: Hy4 is not launched as a standalone technology demonstration, but as the model engine in a commercial agent strategy aimed at enterprise customers.

Access — and the open question

Here lies one of the launch's oddities. The Hy4 Preview in August was, according to Startup Fortune, released under the Apache 2.0 license with downloadable weights. The flagship version, however, arrived only via API: "Anyone hoping to run it themselves can't — not yet," Startup Fortune writes, stating that Tencent has said model weights and system cards will be released later. One secondary account of Tencent's statement says the weights and system cards will be published "within the coming days". Reuters, in its coverage, refers to Hy4 as an "open source" model — a move that should be read as the news agency's rendering of Tencent's positioning, not as a documented status. As of publication time, neither the license nor the timing of the weights has been confirmed.

In practice, this means Hy4 at launch is a service, not an artifact that can be inspected. The free quota runs November 4–15, and thereafter the prices reported above apply — with the noted uncertainty about which figures are correct.

How to weigh this

Bloomberg describes Hy4 as a major advance for Tencent in the pursuit of the AI pioneers, but at the same time as a system that "still lags the best proprietary models", and the team is said to have acknowledged a tangible gap to OpenAI and Anthropic. The reading that best withstands the evidence: Hy4 is a real capability extension — vision, computer use, multi-application workflows — but how well it actually measures up against competitors, who to trust on the hallucination and pricing figures, and whether the weights actually arrive with an open license are questions none of the available sources can answer today. The timing of when the weights and system cards actually become available is thus the most important item to watch in the coming days.

Sources:

• WinBuzzer: Tencent Unveils Flagship Hy4 With Computer-Use Capabilities for Professional Workflows
• Bloomberg (via Yahoo Finance): Tencent Declares Latest AI Model Ranks Among China's Best
• Reuters (via MSN): Tencent rolls out AI model trained with limited cybersecurity data amid hacking fears
• Startup Fortune: Tencent launches Hunyuan Hy4 with vision and the ability to click through apps
• MSN: China's Tencent Hy4 rivals world's top frontier models with 7.2% hallucination rate