Tencent's HunyuanImage 3.0 Beats GPT-Image as the Largest Open-Source Image Model
Tencent open-sources HunyuanImage 3.0, an 80B MoE model that reasons before generating images, rivaling closed-source leaders like GPT-Image and Midjourney.
Takeaways
- Tencent open-sourced HunyuanImage 3.0, the largest open-source image generation model at 80B parameters (13B active per token via MoE).
- Uses a unified autoregressive LLM backbone (Hunyuan-A13B) instead of a standard diffusion transformer, enabling native Chain-of-Thought reasoning before image generation.
- Three checkpoints released: base T2I, Instruct (with image editing and CoT), and a distilled version needing only 8 sampling steps.
- Outperforms GPT-Image, Seedream 4.0, and Nano Banana in head-to-head human preference (GSB) evaluations.
- Supports multi-image fusion (up to 3 inputs), instruction-driven editing, 1,000+ character prompts, and bilingual Chinese/English generation.
- Hardware requirements are steep: base model needs 3×80GB GPUs; Instruct variants need 8×80GB GPUs. Commercial use is permitted under the Hunyuan Community License.
Tencent has open-sourced HunyuanImage 3.0, a native multimodal image generation model that takes a fundamentally different approach from every major open-source image generator before it. Instead of the standard diffusion transformer (DiT) pipeline, it is built on top of a Mixture-of-Experts large language model and reasons through your prompt before generating a single pixel. The result is the largest open-source image generation model ever released, and it claims performance on par with closed-source commercial systems like GPT-Image and Seedream 4.0.
Not just another diffusion model
Unlike common diffusion-transformer image generators, HunyuanImage 3.0 uses a unified autoregressive native multimodal architecture that combines multimodal understanding and image generation in a single framework. The backbone is Hunyuan-A13B, Tencent's own decoder-only LLM. It employs an MoE configuration comprising 64 experts, with 8 experts activated per token, accompanied by one shared MLP, resulting in approximately 13 billion activated parameters per token, balancing expressive capacity with computational efficiency.
The key architectural idea is borrowed from the Transfusion paper: text tokens are generated autoregressively (like a standard LLM), while image tokens are generated via a diffusion process, and both happen inside the same model with the same weights. The model operates in a unified autoregressive framework that natively handles both multimodal understanding and generation. Instead of relying on separate diffusion transformers or cascaded pipelines, it integrates text and image modalities directly, allowing seamless reasoning over interleaved text, dialogue, and visual inputs.

One particularly interesting finding from the technical paper: the MoE experts naturally specialize by modality. Analysis shows MoE experts specialize by modality across layers, optimizing text-image fusion and reducing computation by activating only a subset of the MoE at each inference step. The model didn't need to be explicitly told which experts should handle text versus images; it learned this division on its own.
Chain-of-Thought for pixels
The most distinctive capability is the model's ability to think before it draws. What truly sets HunyuanImage 3.0 apart is its built-in reasoning process. The model employs a native Chain-of-Thought schema during inference, where it internally thinks through the user's intent step-by-step before producing pixels. Give it a sparse prompt like "make a New Year greeting card" and the Instruct variant will expand that into a full visual specification, then generate the image from that expanded description.