Hy3-preview (Reasoning)

Hy3-preview (also known as Hunyuan-3-Preview) is a large-scale Mixture-of-Experts (MoE) language model developed by Tencent, released in April 2026. It serves as the initial public preview of the third-generation Hunyuan series, introducing a "hybrid fast-and-slow thinking" architecture. The model is designed to optimize the trade-off between inference speed and reasoning depth, making it particularly effective for complex agentic workflows and technical problem-solving.

The model utilizes a 295-billion parameter architecture, though it only activates approximately 21 billion parameters per token during inference. This structure includes 192 routed experts and one shared expert, with a top-8 activation strategy. A notable architectural feature is the integration of a 3.8-billion parameter Multi-Token Prediction (MTP) layer, which facilitates speculative decoding to reduce first-token latency and improve overall generation efficiency.

Capabilities and Reasoning

Hy3-preview is distinguished by its native support for configurable reasoning levels. Users can adjust the model's depth of thought via a reasoning_effort parameter (ranging from "none" for instant responses to "high" for deep chain-of-thought). This allows the model to allocate additional compute to verify steps in mathematics, coding, and scientific reasoning tasks. On benchmarks such as SWE-bench Verified and various STEM olympiad exams, the model has demonstrated performance comparable to significantly larger dense models.

In addition to its reasoning capabilities, the model is optimized for agentic tasks and tool-use. It was trained on structured grammar for native tool calling and exhibits high stability in multi-step planning and long-context reasoning. It supports a context window of up to 256K tokens, enabling it to process extensive technical documentation or entire code repositories for development tasks.