Summary

JuZhou 1.0 is an ultra-lightweight, edge-native text-to-image foundation model designed for fully offline, on-device execution, with a focus on native Chinese language understanding. It enables privacy-preserving, fast image synthesis on mobile devices, uniquely trained using exclusively domestic Chinese AI accelerators.

How It Works

The model features a compact 0.387B parameter architecture (0.385B U-Net + 1.90M VAE decoder) optimized for edge deployment. It utilizes Rectified Flow and DMD2 distillation for rapid 4-step inference. Key innovations include native Chinese semantic alignment trained on a 9M Chinese image-text corpus, eliminating external translation, and training entirely on Sugon K100 AI accelerators, validating domestic hardware for large-scale generative AI.

Quick Start & Requirements

Direct code installation is not detailed. Users can experience JuZhou 1.0 via the Mojie Android app.

• App Download: https://www.pgyer.com/mojiemobilellm-android
• Project Page: https://hswai2026.github.io/JuZhouV1/
• Technical Report: https://github.com/Codecode-X/Juzhou/tree/main/JuZhou_Technical_Report Training requires Sugon K100 AI accelerators. Inference is supported on Android (MNN + QNN) and iOS (Core ML).

Highlighted Details

• Achieves a GenEval score of 0.69, outperforming larger models like SDXL (0.55) and SD3-Medium (0.62) with a significantly smaller footprint.
• Enables efficient mobile deployment with inference times around 1.6-4.5 seconds on high-end mobile SoCs.
• Demonstrates superior performance on Chinese poetry-to-image generation and was trained end-to-end on 224 domestic Sugon K100 AI accelerators.

Maintenance & Community

Core contributors include HSW Group and academic partners. Support from Sugon and Hunan Provincial R&D programs is acknowledged. A project page and public launch event indicate active development, with JuZhou V2.0 preparation underway. No specific community channels are listed.

Licensing & Compatibility

No software license is specified in the README. Compatibility is demonstrated for Android and iOS mobile platforms using MNN, QNN, and Core ML. Training infrastructure is domestic Chinese hardware.

Limitations & Caveats

Source code and model weights are currently unavailable pending internal review and compliance clearance. Training requires specialized, non-NVIDIA domestic AI accelerators (Sugon K100).