Tencent today launched and open‑sourced its Hy3 preview model. Hy3 preview is the most intelligent model in the Hy series to date, and it has proven to deliver best‑in‑class performance among models of the same size for context learning, complex reasoning and conversational abilities.
Hy3 preview uses a Mixture-of-Experts (MoE) architecture with a total of 295 billion parameters and 21 billion activated parameters, supporting contextual reasoning up to 256K tokens. Tested and integrated within Tencent’s core AI products such as Yuanbao, CodeBuddy, WorkBuddy and ima, Hy3 preview is grounded in real-world scenarios, and has demonstrated significantly better performance than previous Hy iterations in daily conversations, coding and AI Agent-led tasks.
Yao Shunyu, Tencent’s Chief AI Scientist, stated that the Hy3 preview marks the first step in rebuilding the Hunyuan large model. Through this open-source release, the team hopes to gather authentic feedback from the open-source community and users to improve the practicality of the official Hy3 version.
At the same time, they are continuing to scale up pre-training and reinforcement learning to raise the model’s intelligence ceiling. By deeply co-designing with Tencent’s wide range of products, they aim to steadily enhance the model’s real-world performance across scenarios, while also beginning to explore distinctive model capabilities.
Technical optimizations during the pre-training of Hy3 resulted in an over 40% performance uplift and substantially improved training stability. During the reinforcement learning stage, Hy3 preview addressed stability issues associated with large-scale and long-duration training, implemented a unified agent framework and enabled multiple types of integrated training.
Hy3 preview tops benchmarks in context learning and advanced reasoning
Multiple evaluation results have shown Hy3 preview excelling in context learning and reasoning, enabling the model to understand complex instructions and long texts to execute real-world tasks.
- On mainstream instruction-following and long-text test sets, Hy3 preview outperforms leading domestic models and achieves scores very close to leading large international models on global benchmarks.
- For STEM applications, Hy3 preview performed exceptionally on high-difficulty benchmarks, and has successfully passed international PhD Qualifying Exams.
- Hy3 preview is also proving to be one of the most practical model choices for complex agent scenarios such as OpenClaw. It equals or outperforms same-sized models for agent-led tasks in coding (performed within development environments) and search executions (retrieving, filtering and integrating information from open sources).
Hy3 preview available on open-source platforms and Tencent Cloud
Hy3 preview is available as an open-source model on GitHub and Hugging Face, enabling the developer community to fine-tune the model to create specialized large-model applications tailored to specific business needs.
On Tencent Cloud, Hy3 preview offers cost-effective pricing, with input costs starting at approximately USD 0.18 per million tokens and output costs starting at approximately USD 0.59 per million tokens. In addition, Tencent Cloud has introduced a customized Hy3 preview Token Plan package, with personal plans starting at around USD 4.10 per month to provide an accessible option for developers building agents and next‑generation applications.
Hy3 preview has also officially been listed on global mainstream API aggregation platforms such as OpenRouter, as well as on agent development kits such as OpenClaw, enabling “plug-and-play” workflow integration.
Hy3 preview Experiential Website: OpenRouter(free access for a limited two-week period):
https://openrouter.ai/tencent/hy3-preview:free
