Microsoft has unveiled a new lightweight language model called Phi-3 mini, which is specifically designed to run on modern smartphones and offer performance comparable to OpenAI’s GPT-3.5. This updated version of Microsoft’s language model has been trained with 3.3 billion tokens from larger and more advanced datasets compared to its predecessor, Phi-2, which was trained with 1.4 billion tokens.
Phi-3 mini has 3.8 billion parameters and can fit on a modern smartphone, taking up approximately 1.8GB of memory and can be quantified to 4 bits. The researchers tested the model on an iPhone 14 with an A16 Bionic chip, where it ran natively and offline at a speed of over 12 tokens per second. The overall performance of this model is comparable to larger models like Mixtral 8x7B and GPT-3.5, making it ideal for use in conversational chat formats.
Phi-3 mini uses a transformer decoder architecture that supports 4K text length and is based on a block structure similar to Meta’s Llama 2, supporting packages developed for Llama 2. This makes it easy for developers to integrate the model into their applications and take advantage of its capabilities.
In addition to Phi-3 mini, Microsoft has also trained two other models in the same family: Phi-3 medium with 14 billion parameters and Phi-3 small with 7 billion parameters, both trained with 4.8 billion tokens. The company is focused on providing efficient and secure language models for a variety of applications, aligning with their values of robustness and security.