Deepin 25 makes breakthrough with risc-v chip support for offline AI inference

On September 17, 2026, Deepin 25 has successfully adapted to the Jindie Space K3 RISC-V chip, enabling offline large model inference on desktop systems. This achievement marks the first time a RISC-V-based operating system can run large models without internet access. The K3 chip integrates three core types, including 8 A100 AI cores with 60 TOPS computing power, capable of processing 30 billion parameter models at over 10 tokens per second. Deepin 25 also includes optimizations for the RVA23 specification, enhancing compatibility with the chip's architecture. The system now supports local AI models like Xiao Wu, leveraging the SpacemiT matrix operation extension for efficient task scheduling. The latest RISC-V experience image for K3 is available in the deepin-ports repository, allowing developers to test and deploy the system. This milestone strengthens the domestic open source ecosystem and paves the way for future cross-major version migrations.