On September 5, the official WeChat account of Mianbi Intelligent released a blog post announcing the launch of the open-source MiniCPM3-4B AI model, declaring “the era of edge-side ChatGPT has arrived.”
The MiniCPM3-4B is the third generation of the MiniCPM series and boasts overall performance that surpasses both Phi-3.5-mini-Instruct and GPT-3.5-Turbo-0125, competing closely with several AI models containing between 7 billion and 9 billion parameters.
Compared to its predecessors, MiniCPM1.0 and MiniCPM2.0, the MiniCPM3-4B offers a more powerful and versatile skill set, making it suitable for a wider range of applications. This new model supports function calling and code interpretation capabilities.
The key differences in the model architecture among the three versions are as follows:
– Vocabulary Size: 123K -> 73K -> 73K
– Number of Layers: 40 -> 52 -> 62
– Hidden Layer Nodes: 2304 -> 1536 -> 2560
– Maximum Length: 4K -> 4K -> 32K
– System Prompt Support: Not supported -> Not supported -> Supported
– Tool Calling and Code Interpreter: Not supported -> Not supported -> Supported
MiniCPM3-4B features a 32K context window and utilizes LLMxMapReduce, allowing it to handle theoretically unlimited context without excessive memory consumption.
Additionally, Mianbi Intelligent has also introduced the RAG suite, which includes the MiniCPM-Embedding model and the MiniCPM-Reranker model. They have released a fine-tuned version called MiniCPM3-RAG-LoRA, specifically for RAG scenarios.