HunyuanImage 3.0

What it is

HunyuanImage 3.0 is an open-source image generation model from Tencent. Its inference code, model weights and technical report were all released on 28 September 2025.^[1] Unlike many image generators built on a diffusion transformer (DiT) backbone, it is a native multimodal model: image understanding and image generation are handled inside one autoregressive framework.^[1] In practice this means the model produces an image the way a language model produces text — one token after another — rather than by iteratively denoising a noisy latent.

Architecture and training

The model is a mixture-of-experts (MoE) network with 64 experts, more than 80 billion total parameters, and 13 billion parameters activated per token.^[1][2] Sparsity of this kind lets a model hold a great deal of capacity while keeping the compute spent on each token well below what the raw parameter count would suggest. Tencent described it as the largest open-source image-generation MoE model, and as "the largest and most powerful open-source image generative model to date".^[1][2]

The technical report covers the pipeline behind the model: data curation, a native chain-of-thought schema, progressive pre-training and post-training.^[2] The chain-of-thought part refers to the model writing out intermediate reasoning steps as part of generating, instead of jumping straight to the final image tokens; progressive pre-training means the pre-training stage is carried out in stages rather than all at once.

Evaluation

Tencent reported two kinds of evaluation.^[1] The human evaluation used Good/Same/Bad comparisons across 1,000 prompts and involved more than 100 professional evaluators. The machine evaluation used a method abbreviated SSAE, which scored 3,500 key points across 12 categories.^[1]

Running it

The base text-to-image model is recommended to run on at least 3×80 GB of GPU memory.^[1] The repository requires Python 3.12 or later and CUDA 12.8, with PyTorch 2.8.0 recommended.^[1] Support for vLLM acceleration was added on 30 October 2025.^[1] The repository also notes that FlashInfer can make inference up to three times faster, with the caveat that the first run may take about ten minutes for kernel compilation.^[1]

The Instruct version

On 26 January 2026 Tencent released HunyuanImage-3.0-Instruct, which adds reasoning, image editing and multi-image fusion, together with a distilled version that generates with 8 sampling steps.^[1] The Instruct version can fuse up to 3 input images in a single generation.^[1] For the Instruct and Instruct-Distil versions, the repository recommends at least 8×80 GB of GPU memory.^[1]

Where it sits

Releasing code, weights and a technical report together means outside researchers and developers can inspect, reproduce and run the model, not just read reported numbers.^[1] The line of releases also shows the same autoregressive MoE framework moving from a text-to-image base model to an Instruct model with reasoning, editing and multi-image input, plus a distilled variant that needs fewer sampling steps.^[1] Together with its open weights and its parameter scale, that is what the "largest open-source" framing rests on.^[1][2]