Tencent releases Hy3 under Apache 2.0 and integrates it across its products

Background

Tencent had first open-sourced a preview of the Hy3 model in April 2026. The July launch was the final release of the model, and it was the first major launch under Yao Shunyu, who took over Tencent's LLM work.^[3]

What happened

Tencent launched Hy3 on 6 July 2026. It is a reasoning model that combines fast and slow thinking.^[1][3] Its architecture is a mixture-of-experts (MoE) design with 295 billion total and 21 billion activated parameters, and its context extends to 256K tokens.^[1]

Hy3 was released under the Apache License 2.0.^[2] Apache 2.0 is a permissive open-source license that allows use, modification and redistribution, so developers and companies can build on the model under its terms.

Key details

Hy3 was integrated into WorkBuddy and CodeBuddy, Yuanbao, Marvis and ima.^[1] WorkBuddy and CodeBuddy are Tencent products that appear in the release's integration list alongside the consumer apps.

On Tencent Cloud TokenHub, API pricing was RMB 1 per million input tokens and RMB 4 per million output tokens.^[1] For a general reader, the price per token is the amount charged for each unit of text the model reads or writes.

Reported scores included 78 on SWE-bench Verified and 90.4 on GPQA Diamond.^[2] SWE-bench Verified tests whether a model can resolve real software issues, and GPQA Diamond is a set of difficult graduate-level science questions.

Why it matters

The launch combined three things in one release: an open-source license, a model integrated into several Tencent products, and a published API price. KrASIA described Hy3 as Yao Shunyu's first major launch since he took over Tencent's LLM work.^[3] A reasoning model combines a fast mode for direct answers with a slow mode in which it works through intermediate steps, so one system can handle quick questions and harder problems. For developers, an open license and a published price are the two details needed to plan use: the license sets the terms for using the weights, and the price sets the expected cost per call.