On April 23, 2026, Tencent Hunyuan introduced the preview version of the Hy3 language model and declared that it would be open-source. This innovative model, known as a hybrid expert model, combines the strengths of fast and slow thinking. It boasts an impressive 295 billion parameters, with 21 billion actively engaged during operation, and can handle a maximum context length of 256K. As the inaugural model developed post-Hunyuan's reconstruction, the Hy3 preview showcases notable enhancements in complex reasoning, instruction adherence, contextual learning, coding proficiency, agent capabilities, and overall reasoning performance. Tencent has proudly labeled it as the most intelligent and cost-efficient Hunyuan model yet.

The Hy3 preview has made its debut across a range of products, encompassing Tencent Cloud, Yuanbao, and ima. It also facilitates seamless integration with open-source agent products such as OpenClaw. Regarding pricing, on Tencent Cloud's large model service platform, TokenHub, the Hy3 preview is competitively priced at a minimum of 1.2 yuan per million tokens for input. For inputs that benefit from cache hits, the cost drops to a minimum of 0.4 yuan per million tokens. Outputs are priced at a minimum of 4 yuan per million tokens. Furthermore, Tencent Cloud has collaborated with Hunyuan to offer a tailored Hy3 preview Token Plan package, with the personal version available for a minimum monthly fee of 28 yuan.