Tencent: Hy3 preview

Tencent: Hy3 preview is a text model from Tencent, released by the provider on 2026-04-22. It sits in the text category and is described by Tencent as a high-efficiency Mixture-of-Experts model built for agentic workflows and production use. The "preview" label in the name signals that this is an early version of the Hy3 line rather than a final, long-term release, which is worth keeping in mind if you plan to depend on it in a shipped product.

The model takes text in and returns text out, so there is no image, audio or file input to plan around. Its context window is 262144 tokens, and the maximum output is 235929 tokens, which means you can feed it very large inputs and still leave room for a long response. A notable part of its design is configurable reasoning: Tencent describes reasoning levels across disabled, low and high modes, so you can dial the amount of internal reasoning up or down depending on the task rather than accepting a single fixed behaviour.

Those traits point to a fairly specific set of uses. The long context suits work where you need to hold a large document set, a long conversation history or a big codebase in view at once. The agentic focus and tool support make it a reasonable candidate for assistants that call functions, chain steps or operate over multiple turns. Configurable reasoning helps in cases where you want fast, direct answers for simple requests and deeper deliberation for harder ones, without switching models. The seed parameter is useful when you need reproducible sampling during testing.

You call it through api14's OpenAI-compatible API using the model id "tencent/hy3-preview". Because the interface follows the OpenAI chat completions shape, existing OpenAI SDKs and clients generally work with only a base URL and key change. The supported request parameters are include_reasoning, max_tokens, reasoning, reasoning_effort, seed, temperature, tool_choice, tools and top_p. The reasoning-related parameters (reasoning, reasoning_effort, include_reasoning) are the ones that map onto the model's configurable reasoning levels, while tools and tool_choice cover function calling and temperature, top_p, seed and max_tokens cover standard sampling and length control.

Since this is a preview release, treat stability and long-term availability as open questions, and check the live price table on this page for current rates before committing. If your workload is not text based, or if you need a model with a settled release history rather than a preview, another model may fit better. For text tasks that benefit from a large context, tool use and adjustable reasoning, it is a straightforward option to evaluate.