HunyuanImage 3.0 and CogView3

last edited

HunyuanImage 3.0 and CogView3 are open-source image model families from Tencent and THUDM. Tencent launched the first model in September 2025. THUDM's line has since moved on to a newer generation, so the older repository is a legacy reference.

HunyuanImage 3.0

Tencent's launch post on X is dated September 28, 2025. It describes a native multimodal model with 80B total parameters. It activates 13B parameters per token. Text and image generation share one autoregressive architecture.

The technical report on arXiv has a third version. That revision is dated June 26, 2026.

Instruct release

Tencent released the Instruct models on January 26, 2026, per its Instruct release post. Instruct targets image editing and multi-image fusion. A distilled 8-step variant shipped alongside it.

The Pandaily report on Hunyuan Instruct lists different size figures than the launch post. The model card on the hosting hub is the reference for sizes.

CogView lineage

THUDM published the older model and a 3B Plus variant in its public repository. That model is no longer the lead of the lineage.

The newer model is a 6B system with native Chinese prompt support. It was released on March 4, 2025. A control variant followed on March 24, 2025. The CogView4 repository README documents both.

The diffusers library offers a CogView4 pipeline in diffusers, documented against stable v0.38.0.

Related pages

The family belongs to AI Art Models — Organized Reference. Related context appears in Open Source Tool Sustainability. Video coverage includes Open Source Video Generation Landscape and a second snapshot.

Sources