First Open-Source Image Generation Model Trained Fully on Domestic AI Chips
This article examines an open-source multimodal image generation model whose entire pipeline, from data preprocessing to large-scale pretraining, ran on domestic AI chips. It covers the training toolchain and operator-level optimizations, the hybrid autoregressive and diffusion architecture behind its text rendering, support for non-standard resolutions, and the model's open-source release and API pricing.
From Compute Restrictions to Domestic Breakthrough: An Industry Signal
Two recent developments are worth considering together. They appear unrelated, but they reflect a subtle shift in the competitive picture for large models. First, export restrictions on certain high-end chips from major international GPU vendors have reportedly loosened. Second, a leading domestic AI company and a major technology company jointly open-sourced an image generation model. Its entire process, from data preprocessing to large-scale pretraining, was completed on domestic AI chips.
Comparing the two points to a deeper trend. Even if overseas supply of high-end compute improves, China’s push toward technological self-reliance has not stalled. Under this pressure, it has produced notable results. This article focuses on the second development and breaks down the technical details and engineering practices behind this open-source multimodal model, which is a milestone for the field.
Full-Stack Domestic Training: Not Just Running, but Running Well
For a long time, the industry held a common view: domestic AI chips are adequate for inference, but large-model training demands extreme parallel efficiency and stability, and domestic chips always seemed to fall short. This release challenges that view.
The training is considered a breakthrough mainly because it achieves full-pipeline domestic adaptation, not simple porting. The engineering approach includes the following steps:
In-house training suite covering the full lifecycle: Built on a domestic open-source deep learning framework, the R&D team developed a complete training toolchain. It links data preprocessing, pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL) based on human feedback. The suite does not simply call existing frameworks. It includes substantial customization at the operator and scheduling layers to match the architecture of domestic chips.
Dynamic-graph multi-level pipelining to remove dispatch bottlenecks: Domestic accelerator cards have specific characteristics in instruction dispatch efficiency. To address them, the team adopted dynamic-graph multi-level pipelining. This technique splits the computation graph into fine-grained segments. Multi-level buffering and asynchronous dispatch significantly reduce idle time for compute units, which keeps hardware highly utilized.
Multi-stream parallelism to overlap communication and computation: In distributed training, communication overhead often limits scalability. The team used a multi-stream parallel strategy. Data can move between devices while compute units run other tensor operations, so communication is hidden behind computation. At the scale of a thousand-card cluster, this optimization can deliver near-linear or even super-linear efficiency gains.
Operator-level tuning to tap native chip performance: For low-level operators, the team did not rely on general-purpose operator libraries. Instead, it rewrote or fused key operators for the architecture of domestic chips. These include the AdamW EMA optimizer (adaptive moment estimation with weight decay), computation-communication co-optimization (COC), and RMS Norm. This approach of optimizing down to the level of circuit design logic is key to overall training efficiency.
Together, these optimizations allow the model to train on domestic compute clusters with efficiency and stability that can support frontier research. For the compute-heavy stages of large-model development, this adds a secure, controllable option with solid performance.
Hybrid Architecture: Letting the Model Understand and Render Accurately
The model’s strong performance in text rendering, especially Chinese typesetting, comes from its architecture.
It does not use the currently popular pure diffusion or pure autoregressive structure. Instead, it returns to a division-of-labor approach with a hybrid architecture: an autoregressive language model combined with a diffusion decoder.
Autoregressive backbone (responsible for understanding): The large autoregressive component interprets complex semantics, parses user instructions, and plans the global layout. It is essentially a strong vision-language model. It converts compound instructions, such as “a poster with the main title ‘New Product Launch’, a tech-blue gradient background, and a product image in the lower right corner,” into a structured intermediate representation.
Diffusion decoder (responsible for rendering accurately): This decoder is based on the diffusion transformer (DiT) architecture and includes a dedicated text encoder. Semantic vectors from the autoregressive component guide denoising sampling in latent space. The text encoder provides precise character-level glyph features, so text in the generated image avoids stroke errors, missing components, and mismatched meaning.
This design aligns closely with the “cognitive generation” direction that the industry has promoted in recent years. The core idea is that image generation should combine knowledge reasoning with visual presentation, rather than only fitting a pixel probability distribution. A language model handles “thinking” and planning, and a diffusion model handles “execution” and rendering. This combination gives the model notably higher accuracy than traditional methods on tasks involving dense text and complex logical relationships.
Long-Text Rendering and Arbitrary Resolution: Addressing Design Pain Points
For designers and content creators, the most frustrating problems with text-to-image tools are garbled text and limited output sizes. The model addresses both.
Complex visual text generation: The model ranked first among open-source models on two public benchmarks. This result comes mainly from strong supervision of the text portions of image-text training data, along with the advantages of the hybrid architecture described above. By separating semantic understanding from glyph rendering, the model handles long text, multiple paragraphs, and mixed typography. It can directly generate posters or social media images that contain large blocks of Chinese copy.
Non-standard resolution support: Most image generation models work with a few fixed aspect ratios. When a user needs a poster ratio such as 1240x1600, the output must be cropped or stretched afterward, which unbalances the composition.
