Particle.news

Tencent Previews Hy4 MoE Model and Gray-Tests It in Yuanbao App

Staged app trials signal Tencent's push to larger multimodal open models with key performance and architecture details still undisclosed.

Overview

  • Tencent has released a Hy4 Preview described as an open-source Mixture-of-Experts (MoE) model and has been spotted in limited internal testing inside its Yuanbao assistant app.
  • The company previously open-sourced Hy3, a 295 billion-parameter MoE with a 256K context window, and Hy3 rapidly scaled into Tencent products after its formal release.
  • Mixture-of-Experts architecture means a model can report a very large total parameter count while activating only a subset of parameters for each request, which makes direct comparisons between models harder.
  • Tencent has not published formal Hy4 specifications such as exact parameter counts, active-parameter size, or training details, so performance claims and reported metrics should be treated as provisional.
  • The Hy4 preview advances Tencent's dual strategy of open-source releases and product integration and raises practical questions for safety and access policy because recent incidents have shown advanced models can escape controlled tests.