MiniCPM3-4B
| Property | Value |
|---|---|
| Model Size | 4B parameters |
| License | Apache-2.0 |
| Languages | English, Chinese |
| Context Window | 32k tokens |
| Paper | arXiv:2404.06395 |
What is MiniCPM3-4B?
MiniCPM3-4B is the third generation of the MiniCPM series, representing a significant advancement in compact language models. Despite its relatively small size, it demonstrates performance comparable to or exceeding many 7B-9B models, including GPT-3.5-Turbo-0125. The model excels in both English and Chinese language tasks, featuring advanced capabilities like function calling and code interpretation.
Implementation Details
Built on the Transformer architecture, MiniCPM3-4B incorporates several innovative features that enable its impressive performance. The model supports bfloat16 precision and can be deployed using both the Transformers library and vLLM for optimized inference.
- 32k context window with LLMxMapReduce for theoretically infinite context handling
- Built-in support for function calling and code interpretation
- Optimized for both CPU and GPU deployment
- Comprehensive chat template implementation
Core Capabilities
- Strong performance in multilingual tasks (MMLU: 67.2%, CMMLU: 73.3%)
- Advanced mathematical reasoning (GSM8K: 81.1%, MathBench: 65.6%)
- Robust code generation (HumanEval+: 68.3%)
- Superior function calling abilities (BFCL v2: 76.0%)
- Competitive performance in general benchmarks (MT-Bench: 8.41)
Frequently Asked Questions
Q: What makes this model unique?
MiniCPM3-4B stands out for achieving high performance with a relatively small parameter count, making it more accessible for deployment while maintaining competitive capabilities with larger models. Its balanced performance across multiple domains and languages makes it particularly versatile.
Q: What are the recommended use cases?
The model is well-suited for a wide range of applications including multilingual text generation, mathematical problem-solving, code generation, and function calling tasks. It's particularly effective for applications requiring balanced performance across English and Chinese languages while maintaining reasonable computational requirements.
