OpenBMB releases open-source MiniCPM-V 4.7 vision models for on-device use
Article timeline
The articles in this story. Times are ET.
OpenBMB@OpenBMBOfficialOpenBMB releases MiniCPM-V 4.7 with Self-Replay Fine-Tuning and multi-device deploymentAIOpenBMB announces MiniCPM-V 4.7, which it says scales from mobile devices to PCs and edge systems. The post introduces Self-Replay Fine-Tuning, a method it describes as a low-cost way to customize models while preserving general performance. It says Day-0 hardware support covers Apple, Intel, AMD, Arm and Rockchip, with more platforms reached through FlagOS, and that the full stack is open source.
- OpenBMB@OpenBMBOfficialMiniCPM-V 4.7 released with benchmark results
AIOpenBMB publishes performance results for MiniCPM-V 4.7 in a post that links to the full benchmark table. The post gives no scores or other figures in its text.
- OpenBMB@OpenBMBOfficialPickOpenBMB releases MiniCPM-V 4.7, open-source multimodal models for on-device use
AIOpenBMB announces MiniCPM-V 4.7, a new open-source multimodal model series with a 35B-A3B model and a 1B model. The 35B-A3B model uses LES to reach 17+ tokens/s on iPad Pro M4 (16GB) and 13+ tokens/s on iPhone 17 Pro Max, using 3GB of memory. The post also reports a 5× speedup on iPad Pro M4 and 90.4 on OCRBench at 16× compression, 5.2 points ahead of Qwen3.6 35B-A3B.
- OpenBMB@OpenBMBOfficialOpenBMB releases MiniCPM-V 4.7 with Self-Replay Fine-Tuning and cross-device deployment
AIOpenBMB announces MiniCPM-V 4.7, an open-source model it says can scale from mobile devices to PCs and edge systems. It introduces Self-Replay Fine-Tuning, which the company says customizes models while preserving general performance, and lists day-0 hardware support for Apple, Intel, AMD, Arm and Rockchip, with more platforms via FlagOS. The post also names supported inference, fine-tuning and quantization tools and links to GitHub, Hugging Face, ModelScope and a web demo.