@AdinaYakup: Tencent just released HY3 - 295B / 21B MoE , 256K context - Apache 2.0 - FP8 version included - Switchable reasoning: n…

X AI KOLs Following Models

Summary

Tencent released HY3, a 295B/21B parameter mixture-of-experts model with 256K context length, Apache 2.0 license, FP8 quantization, and switchable reasoning levels, reducing hallucinations by half.

Tencent just released HY3 🔥 - 295B / 21B MoE , 256K context - Apache 2.0 - FP8 version included 👀 - Switchable reasoning: no think / low / high - Hallucinates half as often as before - Stable across agent frameworks https://t.co/yJD4AekKTF
Original Article
View Cached Full Text

Cached at: 07/06/26, 12:06 PM

Tencent just released HY3 🔥

  • 295B / 21B MoE , 256K context
  • Apache 2.0
  • FP8 version included 👀
  • Switchable reasoning: no think / low / high
  • Hallucinates half as often as before
  • Stable across agent frameworks https://t.co/yJD4AekKTF

Similar Articles

tencent/Hy3

Simon Willison's Blog

Tencent released Hy3, a 295B-parameter Mixture-of-Experts model with 21B active parameters, under Apache 2.0 license, outperforming similar-size models and rivaling larger open-source models with 2-5x parameters.

Hy3 (1 minute read)

TLDR AI

Tencent released Hy3, a 295B-parameter MoE model with 21B active parameters, outperforming similar-sized models and rivaling larger open-source models. It is Apache 2.0 licensed, available on Hugging Face and free on OpenRouter until July 21st.