HunyuanImage 3.0
HunyuanImage 3.0: an open 80B+ parameter mixture-of-experts autoregressive model (13B active per token) unifying image understanding and generation.
Native chain-of-thought inside an autoregressive multimodal backbone instead of a diffusion-only pipeline, with staged pretraining and heavy post-training. Code and weights released through Tencent-Hunyuan on GitHub.
- Date
- Sunday, 28 September 2025
- Lab
- Tencent
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Parameters | >80B total, 13B active mixture of experts | authors |
Date is arXiv v1; v-final dated 2026-06-26. Licence terms not checked.
Sources
This record was checked against its sources on 6 October 2026. How we check