Hunyuan Image 3.0

• 🏢 Tencent's native multimodal text-to-image model in the Hunyuan family.
• 📏 80B total parameters with roughly 13B activated via Mixture-of-Experts.
• 🧠 Unified autoregressive framework instead of the prevalent DiT design.
• 👁️ World-knowledge reasoning elaborates sparse prompts into richer scenes.
• 🆕 Instruct variant adds reasoning and image-to-image creative editing.
• ⚡ Distilled checkpoint enables efficient roughly 8-step sampling.
• 🎯 Photorealistic output with strong prompt adherence and fine detail.
• 📚 Handles long, detailed prompts spanning complex multi-subject scenes.

Tencent

Tencent is a Chinese multinational technology conglomerate headquartered in Shenzhen. One of the highest-grossing multimedia companies in the world by revenue, Tencent has built a vast portfolio spanning gaming, social media, fintech, and cloud services. In…

Hunyuan Image 3.0 is Tencent's text-to-image generator and the latest entry in the company's Hunyuan image line. Tencent describes it as a native multimodal model that unifies multimodal understanding and generation within a single autoregressive framework, with the image-generation module released openly. The architecture pairs a Mixture-of-Experts design with roughly 80 billion total parameters and about 13 billion activated during inference, using a Transfusion-style approach to bind text and image tokens.

The most concrete change from its same-family predecessor is structural. According to the technical report, version 3.0 moves beyond the prevalent DiT-based architectures to a unified autoregressive framework that models text and image modalities more directly, which the team links to more contextually rich generation. It also adds world-knowledge reasoning, automatically expanding sparse prompts with contextually appropriate detail.

On the feature side, the model emphasizes photorealistic imagery, strong prompt adherence, and fine-grained detail, alongside support for long, detailed prompts spanning multiple subjects and lighting parameters. The arXiv technical report details the data curation and post-training reinforcement learning behind these behaviors.

Tencent has since shipped additional checkpoints: an Instruct release adding reasoning-based prompt enhancement and image-to-image editing, plus a distilled variant tuned for efficient deployment with roughly 8-step sampling. Weights and code are available through Tencent's Hugging Face repository for self-hosting.