Tencent’s Hunyuan AI: 3D Worlds, Videos, and Business Wins

Tencent keeps pushing its Hunyuan AI models into new territory. A WIRED report shows Riot Games using them to whip up 3D characters and scenes for Valorant—down from a month of work to four options in 60 seconds via text prompts. The same models help with GKART and indie devs. Tencent’s not talking, but insiders say it’s speeding up game prototyping big time.

What It Can Do

Hunyuan handles 3D natively. HunyuanWorld 1.0 spits out interactive scenes, like a Lego valley I tested, per WIRED. Upload video, and a newer version builds 3D from it. Hunyuan 3D makes printable objects, say custom D&D minis.

On video, HunyuanVideo-1.5 packs 8.3 billion parameters and runs on GPUs with 14GB VRAM—no cloud needed. It uses Selective and Sliding Tile Attention for almost 2x faster inference than before, plus a 3D VAE that crushes data 16x spatially and 4x temporally. Output hits 1080p sharpness straight out, with smooth motion.

The foundational Hunyuan got a bump for better coding, math, and science, as Tencent detailed in its Q3 earnings covered by CNBC.

Real-World Uses

Games: Faster content creation at Riot and beyond, cutting design time massively (WIRED).
Business: Boosts ad targeting, player hooks in games, plus efficiency in coding and video production (Tencent CEO Ma Huateng via CNBC).
Cloud: Higher demand for AI services drove revenue up (CNBC).
Creators: Local video gen on consumer rigs like RTX 4090s, open-source for devs skipping subscriptions (WinBuzzer).
OCR: HunyuanOCR nails recognition tasks (Pandaily).
Film/Animation: AI like Hunyuan could produce 30% of long-form content by 2026 (Variety).

This helped Tencent hit 192.9 billion yuan revenue in Q3 2025, beating estimates, with gaming up 16% domestically and 43% internationally. Shares jumped 56.7% year-to-date.

How It Stacks Up

For 3D, Hunyuan leads a key leaderboard, outpacing Microsoft, Meta, Stability AI, and Bytedance models (WIRED). Startups like World Labs’ Marble chase persistent scenes for games or robots. Academic stuff from Stanford and Princeton adds code-based scene gen and AI agents.

HunyuanVideo-1.5 claims strongest open-source video spot, hitting top visual quality and speed. It eyes Meituan’s LongCat-Video locally and nods to OpenAI’s cloud-heavy setups (WinBuzzer). Tencent pairs it with DeepSeek in spots.

China’s AI race heats up—Tencent’s game chops and WeChat integration could give it legs.