① Microsoft claims that the performance of this lightweight AI model is comparable to models such as GPT-3.5, and is also suitable for local use on mobile phones; ② More importantly, this model is much cheaper, which means it can attract a wider customer base through cost advantages.
AFP, April 24 (Editor: Liu Rui) On Tuesday EST, Microsoft launched a lightweight AI model, the Phi3-mini.
Microsoft claims that the performance of this lightweight AI model is comparable to models such as GPT-3.5, and is also suitable for local use on mobile phones. More importantly, the model is much cheaper, which means it can attract a wider customer base through cost advantages.
Microsoft releases small language model
PHI-3-mini is the first of three small language models (SLM) that Microsoft plans to release.
The Phi-3 mini has 3.8 billion measurable parameters. The previous version was the Phi-2 model released by Microsoft Research at the end of 2023. The model has 2.7 billion measurable parameters. Microsoft says the Phi-3 has better performance than the previous version, and its response speed is close to 10 times larger than the model.
Microsoft said that after evaluating the performance of the Phi-3 mini based on academic benchmarks and internal tests, it was found that the model is capable of competing with models such as the “Mixtral 8x7B and GPT-3.5.”
The company said the Phi-3 mini scored 69% on the Massive Multitask Language Understanding (MMLU) test and 8.38 on the Multi-turn (MT) benchmark.
However, Microsoft also acknowledged that although it has capabilities comparable to LLM, the Phi-3 mini is “fundamentally limited by capacity when performing certain tasks.” “The model doesn't have the ability to store too much factual knowledge,” but Microsoft asserts that “this weakness can be solved through search engine enhancements.”
“More than just a little cheaper”
Sassabastien Bubeck, Microsoft's vice president of generative AI research, emphasized: “The Phi-3 is not just slightly cheaper, it's much cheaper. They cost a tenfold difference compared to other models with similar features.”
The small language model is designed to perform simpler tasks, which means it's more suitable for companies with limited resources.
The company said PHI-3-mini will be immediately available on Microsoft's cloud service platform Azure's artificial intelligence model catalog, machine learning model platform Hugs Face, and Ollama, a framework for running models on local machines.
The PHI-3-mini can also be used on Nvidia's software tool Nvidia Inference Microservices (NIM), and is also optimized for its graphics processing unit (GPU).