Phi-3.5

About

The Phi-3.5 family is a suite of three open-source small language models (SLMs), recognized for their high capability and cost-effectiveness. Notable for its utility in multilingual tasks and ability to handle extended context, the Phi-3.5-mini model features 3.8 billion parameters with support for a context length of 128K tokens, making it ideal for tasks such as summarizing lengthy documents. The Phi-3.5-MoE model, leveraging a Mixture-of-Experts (MoE) architecture, exhibits strong performance across multiple languages with 42 billion parameters (6.6 billion active), achieving reduced latency. Equipped with 4.2 billion parameters, the Phi-3.5-vision model is a multimodal system capable of processing both text and images, facilitating functions like optical character recognition and video summarization. All models were trained using high-quality data, including synthetic and filtered content from publicly available websites, and they incorporate safety measures. They are accessible under the MIT license on Hugging Face, offering robust capabilities for developers and researchers 1456.