Hy3-preview (Reasoning)

Released in April 2026, Hy3-preview (also known as Hunyuan-3-preview) is a high-capacity large language model developed by the Tencent Hy Team. It is specifically optimized for complex reasoning, agentic workflows, and STEM-related problem-solving. This model represents a significant architectural shift for Tencent, as it is the first to be trained on their rebuilt infrastructure designed to integrate "fast and slow thinking" capabilities, enabling the model to handle both high-speed conversational tasks and deep, multi-step logical deduction.

The model utilizes a Mixture-of-Experts (MoE) architecture with a total of 295 billion parameters, though only 21 billion parameters are activated during each forward pass. This design provides the reasoning depth of a massive dense model while maintaining the inference efficiency of much smaller systems. A notable architectural feature is the inclusion of a 3.8B Multi-Token Prediction (MTP) layer, which enhances the model's ability to plan and anticipate subsequent tokens, particularly improving performance in coding and mathematical reasoning.

Hy3-preview supports a large context window of 256K tokens, allowing it to process extensive documentation and messy, real-world data for long-context learning. Tencent reports substantial improvements in instruction following and agentic capabilities, with high scores on benchmarks such as SWE-bench Verified for bug fixing and BrowseComp for complex web research. The model is also engineered for extreme cost efficiency, featuring optimizations that reduce end-to-end response times by approximately 47% compared to previous generations.

A unique feature of the Hy3 series is its support for configurable reasoning levels (disabled, low, and high). This allows users to adjust the depth of the model's "thinking" process based on the complexity of the query, effectively trading off latency for deeper reasoning in tasks like academic exams, competitive math, and professional-grade code generation. It has demonstrated strong performance in specialized domains, such as passing the Tsinghua Qiuzhen College Math PhD qualifying exams and the China High School Biology Olympiad.