Tencent brings Hy4 computer use to its apps, with weights still to come

Tencent has pursued enterprise distribution for years. When it introduced Hunyuan in September 2023, it offered Chinese enterprises access through Tencent Cloud APIs and said the model had been connected to 50 of its own products. Those included Tencent Meeting, Tencent Docs, games and financial services. That history matters because Tencent supplies both the underlying model and applications in which customers can use it.

The immediate predecessor was already aimed at coding, office work and scientific research. Tencent’s August 28 preview announcement said that version was open-sourced, with 770 billion total parameters, 49 billion active parameters and a context window exceeding 1 million tokens. Those specifications describe the preview version.

Tencent said it developed the preview’s training data with internal specialists in software engineering, gaming, finance and security, working closely with product teams. That gives it relevant workflows for development and evaluation, while leaving the resulting performance claims subject to external testing.

Visual computer use also has precedent among competing models. The Qwen team’s January 2025 announcement of Qwen2.5-VL, a vision-language successor to Qwen2-VL, described chart analysis and visual agents able to use computers and phones. It offered base and instruction-tuned models in three sizes. Hy4’s launch therefore adds these capabilities to Tencent’s model line within an established category. The announcements do not provide a common test that shows which model is better at completing the same work.

API prices leave task economics unresolved

Standard Hy4 API pricing remains unchanged from the preview. A report on the launch terms lists $0.834 per million input tokens and $2.501 per million output tokens, with a 50% discount on Tencent Cloud TokenHub through December 31. WorkBuddy and CodeBuddy offer a limited free usage quota from November 4 to November 15.

Those are hosted usage prices, not the cost of running the weights on a customer’s infrastructure. Nor do they establish cost per successfully completed task: that depends on how much input and output a workflow consumes, and how many attempts or corrections it requires. The launch reports provide no comparable independent measure of that cost or hardware requirements for the promised release.

Hy4’s standard API prices, before the temporary TokenHub discount. Token rates do not establish cost per successfully completed task. Source: TechNode.

What the weight release needs to resolve

Tencent says it will release Hy4’s weights shortly and publish a system card with fuller safety evaluations, focusing on cybersecurity and biological risks. It says it has worked with the vLLM and SGLang communities and Inferact on support for the open-model ecosystem. It also plans Hy4 flash, promising faster inference and better cost efficiency, according to the launch account.

The company’s own caveat sets a demanding test: it says Hy4 still falls substantially behind leading closed models on difficult internal evaluations of frontier mathematics, long-horizon development, model iteration and visual question answering. Those selected internal tests do not quantify the gap across every use case, but they limit the broader ranking claim.

For developers considering independent deployment, the next decisions depend on the actual weights, deployment requirements and safety evidence. For customers trying the hosted applications now, the unresolved question is whether Hy4 can reliably finish their professional workflows at a competitive total cost. The promised release and comparable external testing are needed to assess both.