Hunyuan Hy4 hands-on: 770B MoE with native vision builds 3D keyboard, theme park and more

量子位 · wechat · 2026-11-04

Tencent's official Hunyuan Hy4 release: 770B total / 49B activated parameters, with first native visual understanding and computer use (CUA), plus upgraded instruction following, long-document and long-horizon agent capabilities. Weights to be open-sourced soon; already live in WorkBuddy and other Tencent products.

QbitAI's hands-on highlights: pixel-accurate reproduction of a retro music player from a screenshot (it even synthesized test tracks via Web Audio and ran 59 headless-Chrome self-tests, all passing); a walkable 3D theme park built from a floor plan; a 16-page product deck where it caught its own layout overflow and validated templates page-by-page before batch-applying; a batch image tool modeled on Squoosh; and mouse-drawing in JSPaint via CUA. Official stats: hallucination rate down from 12.95% to 7.2%, first among open models in a 213-expert blind test of 579 real finance tasks; it aligned 6.99M single-cell mouse embryo sections into a 3D mesh in 1.5 hours with zero human input, and co-authored a counterexample to a math conjecture with Lean-verified proofs with Peking University. Pricing unchanged; a cheaper Hy4flash is coming.

Original post →

Hunyuan Hy4 hands-on: 770B MoE with native vision builds 3D keyboard, theme park and more

More from coding & agent