Rockchip
Rockchip (Ruixin Microelectronics) is a Chinese semiconductor company specializing in integrated circuit design, headquartered in Fuzhou, Fujian. The company's primary products are high-performance application processor SoCs (represented by the RK series) targeting consumer electronics, tablet computers, smart TVs, IoT, and edge AI computing. Rockchip holds a significant position in the domestic embedded processor market, with products widely adopted by device manufacturers globally.
Timeline (last 90 days)
The 2B version of Qwen 3.5 (about 2 billion parameters) has been successfully deployed locally on an RK3588 development board; it can handle text Q&A and single-image Q&A, and can be used through a web interface.
RKLLM is a dedicated large language model inference framework that Rockchip provides for its NPU platform; model weights must be converted through the official toolchain and cannot load GGUF or PyTorch formats directly.
RKNN is Rockchip's NPU inference format for vision models, commonly used for tasks such as classification, detection, and OCR.
MiniCPM-V 4.7 provides day-0 support for Rockchip, Houmo, Intel, AMD, and Arm platforms.
RK3576 是瑞芯微推出的中端 AIoT 处理器，内置约 6 TOPS 算力的 NPU
All Facts (10)
RV1126是瑞芯微专为智能视觉设计的芯片，内置2TOPS算力的NPU
90%Unverified瑞芯微RK3588采用8nm工艺，集成四核A76+四核A55的大小核架构，内置6TOPS算力的NPU，支持8K视频解码
90%Unverified当前国内嵌入式行业大量采用海思、瑞芯微等国产解决方案，尤其在人形机器人、智能安防、工业视觉领域应用广泛
80%Unverified瑞芯微RK3588和联发科天玑系列已将AI推理能力下沉至百元级模组
75%Unverified瑞芯微RK3588芯片具有6 TOPS算力
55%UnverifiedRKNN is Rockchip's NPU inference format for vision models, commonly used for tasks such as classification, detection, and OCR.
50%UnverifiedRKLLM is a dedicated large language model inference framework that Rockchip provides for its NPU platform; model weights must be converted through the official toolchain and cannot load GGUF or PyTorch formats directly.
50%UnverifiedMiniCPM-V 4.7 provides day-0 support for Rockchip, Houmo, Intel, AMD, and Arm platforms.
50%UnverifiedRK3576 是瑞芯微推出的中端 AIoT 处理器，内置约 6 TOPS 算力的 NPU
50%UnverifiedThe 2B version of Qwen 3.5 (about 2 billion parameters) has been successfully deployed locally on an RK3588 development board; it can handle text Q&A and single-image Q&A, and can be used through a web interface.
50%