MiniCPM4 targets edge devices, with optimizations across model architecture, training method, data, and inference system; specific throughput, latency, and comparison results are given in the technical report.
Technical highlights
：MiniCPM4 is pre-trained using 32K long text and length extension is achieved through YaRN technology. In the RULER-NIAH task of finding a needle in a haystack (RULER-NIAH) of 128K long text, it maintained perfect performance at all depths and lengths, demonstrating excellent long context retrieval capabilities.
：On two typical terminal chips, Jetson AGX Orin and RTX 4090, MiniCPM4 is significantly faster than models of the same size in long text processing tasks, and the longer the text, the more obvious the efficiency advantage. On the Jetson AGX Orin platform, the decoding speed of MiniCPM4 is approximately 7 times faster than that of Qwen3-8B.