Tencent Hy3 AI Usage Surges 68-Fold to Claim Top Spot on Global Leaderboard

Tencent Holdings Ltd. expanded its artificial intelligence infrastructure following an abrupt increase in demand that briefly strained its servers after releasing the official version of its Hy3 large language model. Data from tech industry tracking outlet 36Kr shows that total API calls for the new model surged more than 68 times compared to the previous-generation Hy2 version.

Credit: Tencent

The rapid deployment of the text model into corporate and consumer tools triggered a massive consumption of data tokens. Within seven days of its official release, Hy3 secured the number one rank for call volume on the global large-model platform OpenRouter.

Infrastructure strain forced the Shenzhen-based company to inject immediate computing capacity to prevent system outages. This capacity expansion occurred after Tencent integrated the model into WorkBuddy, its proprietary intelligent office assistant.

Tencent also runs the model across other platforms like the desktop automation tool Marvis, the knowledge-base product ima, and its messaging ecosystem.

Chief AI scientist Yao Shunyu oversaw the project from the office of the chief executive officer and president, making this the first major public product rollout since he assumed control of the company's large language model development group. The new model achieves a 90 percent success rate on automated agent tasks across the company's internal software ecosystems.

Tencent released an official statement regarding the processing efficiency of the architecture:

"The model demonstrates significantly enhanced performance relative to models of the same size, while achieving intelligence comparable to flagship models with two to five times its parameter scale."

The underlying setup relies on a Mixture-of-Experts system containing 295 billion total parameters, though it activates only 21 billion parameters per individual request to minimize processing expenses.

It processes a context length of up to 256,000 tokens, enabling corporate users to feed lengthy financial documents, codebase repositories, and game production scripts into the automation pipeline.

Independent testing data from Artificial Analysis and Hugging Face indicates that the model costs 14 cents per one million input tokens and 58 cents per one million output tokens. This pricing configuration positions it among the cheapest options on the international market for high-volume corporate automation.

The company highlighted the commercial performance of the system in its corporate metrics:

"Since the launch of Hy3 preview, its average daily token consumption has increased twenty-fold, reflecting growing market recognition of its positioning as a practical and cost-effective model."

Internal tests from the company state that the model reduced its hallucination rates from 12.5 percent down to 5.4 percent compared to earlier versions.

Common logic and fact errors dropped from 25.4 percent to 12.7 percent, which the tech firm links to a complete rebuilding of its pre-training and reinforcement learning infrastructure started in early 2026.

For specific sub-features, such as the AI PowerPoint generation tool inside Tencent Docs, the update increased successful file creation by 20 percent.

The tech company pushed the software to third-party developer suites including OpenRouter, Cline, Cherry Studio, and Hermes to attract international software engineers. The group published the model under an Apache 2.0 open-source license, allowing outside businesses to host and modify the system on private servers without restrictive fees.

Regarding the future roadmap of the system, the Tencent Hy Team stated:

"Looking ahead, Tencent Hy will continue to accelerate its technical iteration, push the boundaries of model intelligence, and focus on real-world applications, turning leading large model capabilities into accessible productivity for every industry."