Running DeepSeek R1 1.5B Model on VisionFive 2

The on-device deployment of DeepSeek R1 1.5B on the RISC-V-based VisionFive 2-(Video) platform enables edge-native AI agent operation, delivering multi-session conversational intelligence with context-aware reasoning through interactive interfaces.


XinJian® AI Computing Accelerator Card

The XinJian® AI Computing Accelerator Card, developed by Shenzhen Yunji Internet Ecosystem Technology Co., Ltd., is an M.2 2280 form-factor computing module based on the proprietary AXCL heterogeneous computing architecture. Designed for edge-side high-density inference workloads, it enables lightweight AI deployment in scenarios such as smart cities, industrial quality inspection, and educational interaction.

For detailed installation procedures, refer to the AXCL V1.0 documentation:
https://axcl-docs.readthedocs.io/zh-cn/latest/index.html

AX650N

AX650N is a professional smart IP camera SoC with strong AI capabilities. It integrates eight ARM Cortex-A55 processors, a high efficiency NPU that offers 72.0TOPS@INT4 or 18.0TOPS@INT8,an AI enabled ISP pipeline that delivers up to 8K@30fps, and a VPU with H.264/H.265 encoding capabilities. All these functionalities are provided in a low power chip design.

With a high-performance AI-ISP and a power-efficient NPU, AX650N enables various AI innovations and brings an extraordinary AI enabled video and image processing experience to worldwide customers.

2 Likes

Can you tell us a bit more about running DeepSeek on the VisionFive 2?
Or did you use the XinJian AI Computing Accelerator Card?

Yes, in this use case we used the XinJian AI Computing Accelerator Card. However, since the AXCL deb package currently used for VisionFive 2 is an internal beta version provided by AXERA, so we haven’t released detailed documentation/setup guides yet.

2 Likes