88Z AI, 14 October — ModelBest and OpenBMB have jointly released the open-source edge multimodal model MiniCPM-V 4.7, available in 35B-A3B and 1B versions. The 35B-A3B is the world's first open-source MoE multimodal large model capable of lossless, smooth operation on edge devices, requiring only about 3GB of memory and performing well in various benchmarks. The model supports multiple chips and its code is fully open-sourced.
At a glance
- The 35B-A3B is the first open-source MoE multimodal model to run losslessly and smoothly on edge devices, requiring only about 3GB of memory.
- The model supports Day-0 adaptation for multiple chips, with weights, inference, and fine-tuning code fully open-sourced.
The update
ModelBest has partnered with OpenBMB to officially open-source the edge multimodal model MiniCPM-V 4.7, releasing both 35B-A3B and 1B versions. The 35B-A3B model stands out as the world's first open-source MoE multimodal large model capable of lossless and smooth operation on edge devices, demonstrating significant performance advantages. It leads peers in 34 authoritative benchmarks and requires minimal hardware resources, operating with only about 3GB of memory. In practical tests, it achieved a generation speed of over 17 tokens per second on an iPad Pro M4 with 16GB of RAM. Furthermore, MiniCPM-V 4.7 has achieved Day-0 adaptation for multiple chips, ensuring compatibility and efficient performance across various hardware platforms. To foster community development and technology adoption, the team has fully open-sourced the model weights, inference code, and fine-tuning code, allowing developers to freely access and utilize them for future research and application development.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 面壁智能Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.