YT · YOUTUBE// 4h agoMODEL RELEASE

Tencent launches 295B-parameter agentic Hy3 Preview

Tencent has overhauled its AI strategy with the release of Hy3 Preview, a Mixture-of-Experts (MoE) model reconstructed from the ground up for autonomous agentic capabilities. The model features 295 billion total parameters with 21 billion activated, a 256K context window, and is being deeply integrated across Tencent's massive ecosystem of consumer and developer tools including WeChat, QQ, and Tencent Cloud.

// ANALYSIS

Tencent is pivoting from "LLM as a chatbot" to "LLM as an operating system," leveraging its dominant app footprint to embed autonomous agents at scale.

–Reconstructed architecture abandons previous foundations to optimize for reasoning and agentic workflows rather than simple text generation.
–Massive 295B scale with efficient activated parameters allows for industry-leading performance (74.4% on SWE-bench) without prohibitive inference costs.
–Deep integration into ubiquitous platforms like WeChat and QQ provides a unique distribution advantage for "agentic" services that competitors lack.
–Open-sourcing the preview version is a tactical move to foster a developer ecosystem around Hunyuan before the official commercial rollout.

// TAGS

tencenthy3-previewmoeagentic-aillmopen-sourcechinese-ai

DISCOVERED

4h ago

2026-04-26

PUBLISHED

4h ago

2026-04-26

RELEVANCE

8/ 10

AUTHOR

AI Search