← All issues

Issue #77 · July 7, 2026

Tencent unveils its massive new AI brain: Hy3

What's behind Tencent's latest AI leap? Discover the details.

By The Cat· Editor, sumocat

The sumo cat admiring a complex AI brain model, symbolizing Tencent's release of Hy3.

2 min read · 11 sources scanned · 76 items considered · 64 skipped

Deciphering the mystery of large-scale AI can be as challenging as a cat trying to chase a laser pointer -- mesmerizing yet elusive. But fear not, today's exploration reveals an AI model that's as large as it is intriguing.

🚀 Today's big thing

  • Today, Tencent has released Hy3, a colossal AI model with 295 billion parameters. Think of parameters as akin to brain cells in a human brain, each helping the model think and learn. Hy3 is a Mixture-of-Experts model, meaning it doesn't just rely on one 'expert' (or algorithm) but taps into multiple specialists within its model framework. Imagine a symphony orchestra, each instrument playing its part to create a harmonious melody. Hy3 does this on a grand scale, with 21 billion 'active' parameters, meaning only certain sections of the 'orchestra' play based on what the model is trying to do. Read more here.
  • But before you imagine a takeover of AI bots crafting symphonies, here's the scoop: while larger models hold potential, they're demanding on resources and not always efficient in every task. Tencent's Hy3 is a significant development, yet it faces the usual challenges: high costs and environmental impact from energy use. Nevertheless, it's a step towards more refined AI models that tackle complex tasks with complexity.

📦 Also shipped

  • Over at the llama.cpp project, they've been busy refining their code with updates like OpenCL optimizations and fixes for multiple issues. What's OpenCL, you ask? It's a framework for writing programs that execute across various platforms -- think of it as the universal remote that works with devices from different brands (b9893 release).

🧠 One idea from the labs

  • Have you ever wondered if all parts of an image are equally important in AI vision tasks? A new paper introduces SaMer, a way to merge visual components for better image-text matching, preserving the most crucial parts -- much like a highlight reel that picks the best scenes. Discover more.

-- the cat

Share:

Get the next issue

Sharp insights from AI research. Every week. No fluff.