Tencent Hunyuan 混元

What people say

Tencent open-sourced HunyuanVideo in December 2024 at 13 billion parameters, which the company billed at the time as the largest openly-released video generation model, trained specifically to beat closed competitors including Runway Gen-3 and Luma 1.6 on independent evaluations of text alignment, motion quality and visual fidelity. The follow-up, HunyuanVideo-1.5, landed in November 2025 and made the bigger claim: without an obvious quality trade-off, it cut the hardware floor from roughly 45GB of VRAM down to about 14GB, low enough to run 720p, 121-frame generation on a single consumer RTX 4090. Reviewers testing it against Wan 2.2 and LTXVideo-13B in early 2026 consistently flagged its natural motion and physics, cloth simulation, multi-person scenes, as the standout, with the caveat that visual clarity holds up even when motion coherence occasionally breaks down.

HunyuanImage-3.0, released September 2025, is the image half of the family: an 80-billion-parameter native multimodal model built on a mixture-of-experts architecture with 64 expert modules, which briefly took the #1 spot on LMArena's text-to-image leaderboard. Independent write-ups highlight Chinese-character text rendering, historically a weak point for every major image generator, as a specific strength, alongside strong performance on complex, multi-element scene composition where the model's LLM backbone helps it parse long, detailed prompts.

The real constraint isn't quality, it's the license. The Tencent Hunyuan Community License defines its "Territory" as the world excluding the EU, UK and South Korea; using, modifying or distributing the models in those jurisdictions isn't permitted under the standard terms. It also imposes a 100-million-MAU commercial gate (cross that threshold and you need Tencent's explicit sign-off), forbids using model outputs to train or improve a competing AI model, and layers on a 20-category acceptable-use policy covering disinformation, malware, and high-stakes automated decisions in law enforcement or medicine. None of this is unusual for a corporate "open" release, but it's meaningfully more restrictive than a true Apache or MIT license, and worth checking against your jurisdiction before committing engineering time.

Summary of public user & expert reviews, compiled by RECATOOLS.