ATRIUM search → argument graph

DefinitionArticle

Hy3 is a 295B-parameter Mixture-of-Experts model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team.

Tencent describes the technical architecture of its new Hy3 model: a 295B-parameter MoE model with 21B active parameters. ✦ AI generated

Tencent Hy Team · Simon Willison's Weblog · 2026-07-06 · original ↗

Share card ↓Export .md Export .txt

Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team.

Read full article ↗excerpt · fair-use quotation

▦ Tencent Hy Team ✳ Hy3

Related moments

Mechanism◆Article · 160 words

Inkling separates the cost of storing a model from the cost of running it by activating only 6 of 256 experts per token, processing roughly 41 billion of its 975 billion parameters at a time.

Thinking Machines · ByteByteGo Newsletter

Mechanism◆Article · 117 words

The infrastructure reality is that 'open-weight' does not mean easy to run: a 2.4T-class MoE like Qwen3.8-Max is not a commodity local model, requiring massive memory and accelerator counts similar to Kimi K3.

AINews · Latent Space

FactArticle · 48 words

Zyphra, which trains its models on AMD GPUs and is seen as an insider tip in the research community for its detailed architecture-focused tech reports, released new flagship MoE models: a 74B-A4B and an 8B-A0.6B.

the author · Interconnects