Intel / post

Tencent releases Hy3 preview as open-weight agent model

Source: x.com
Date: Published Apr 24, 2026

Context

Tencent used this April 24, 2026 post to say that Hy3 preview, the newest model in its Hy language-model series, could be downloaded from GitHub, Hugging Face, ModelScope and GitCode. A preview is an early public release meant to collect feedback before a finished version: Tencent's own Hy3 preview announcement says the release would help it improve the model "ahead of its official launch." The post itself offers only general claims of strong reasoning, long- understanding and capabilities, meaning the ability to carry out multi-step tasks with tools rather than only answer a prompt.

The size comes from that announcement, not the post. Tencent describes Hy3 preview as a model with 295 billion parameters in total, of which 21 billion are active for each token: the model sends each piece of text through a subset of its sub-networks, so each step needs far less computation than the full size suggests, although anyone hosting it still has to store all the weights. The post's text gives no benchmark results, so its capability claims are Tencent's own. The finished Hy3 followed in July under the Apache 2.0 license, covered in the official Hy3 post.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.

Read the source x.com