A new Chinese text-to-image model called JuZhou 1.0 can generate images directly on your smartphone in about 4.5 seconds, without sending any data to the cloud. The breakthrough challenges the assumption that powerful image generation requires expensive cloud servers, instead packing the technology into a model small enough to run offline on mobile devices.

JuZhou 1.0 achieves this efficiency through a remarkably compact design. The model's core image-generation engine contains just 0.387 billion parameters, roughly one-fifth the size of Stable Diffusion XL (SDXL), which has 2.6 billion parameters. Despite its smaller footprint, JuZhou 1.0 scored 0.70 on GenEval, a standard benchmark for image quality, outperforming SDXL (0.55), Stable Diffusion 3-Medium (0.62), and other published baselines.

The model was trained entirely on Chinese domestic AI accelerators, specifically Sugon K100 clusters, without relying on NVIDIA GPUs. This represents a significant milestone for China's push toward computational independence in artificial intelligence development. The training infrastructure used 56 compute nodes with 224 K100 data center units, demonstrating that domestic hardware can support large-scale generative AI training.

How Does JuZhou 1.0 Achieve Such Speed and Efficiency?

The model uses several technical innovations to compress image generation into just four denoising steps, down from the 28 to 50 steps typical of earlier diffusion models. On a Xiaomi 17 Pro Max smartphone powered by Snapdragon 8 Elite Gen 5, the core denoising process completes in approximately 1.6 seconds, with the full pipeline including on-device prompt refinement taking 4.5 seconds.

• Rectified Flow Training: A technique that optimizes the path from noise to image, reducing the number of steps needed for generation.
• DMD2 Distillation: A compression method that transfers knowledge from larger models into smaller ones, maintaining quality while reducing size.
• Native Chinese Semantic Alignment: The model was trained on 9 million curated Chinese image-text pairs, allowing it to understand Chinese prompts directly without requiring external translation.

Why Does Running AI Locally on Your Phone Matter?

Cloud-based image generation services require users to upload their prompts and reference images to remote servers, creating privacy vulnerabilities. Data can be intercepted during transmission, collected without explicit consent, or exposed in security breaches. JuZhou 1.0 eliminates these risks by processing everything locally on the device. Once installed, the model requires no internet connection and no data leaves the phone.

The researchers also built a specialized version for classical Chinese poetry-to-image generation, trained on 1.77 million poem-image pairs. This demonstrates the model's ability to capture nuanced cultural contexts without relying on external translation modules. The poetry application showcases how lightweight models can be adapted for specialized domains while maintaining the privacy and speed benefits of on-device processing.

JuZhou 1.0 has been deployed across both iOS and Android platforms, with full-stack adaptation completed for major mobile operating systems. The model represents a shift in how the AI industry thinks about image generation: not as a cloud service requiring constant connectivity, but as a tool that can run independently on consumer hardware. This approach could reshape expectations around privacy, latency, and accessibility in generative AI applications.

The development also signals China's broader strategy to reduce dependence on Western semiconductor and software ecosystems. By demonstrating that domestic AI accelerators can train competitive generative models, JuZhou 1.0 provides a reference point for future software-hardware co-design efforts in the region. As edge AI deployment becomes increasingly important for mobile and IoT applications, models like JuZhou 1.0 may influence how other organizations approach the trade-off between model capability and computational efficiency.