Phi-3 was Microsoft Research's push into small language models that punch above their size: Phi-3-mini, at just 3.8 billion parameters, matched the benchmark performance of models roughly twice its size, including rivaling GPT-3.5 and Mixtral 8x7B on several tasks, while being small enough to run on a laptop or even a phone. The lineup also included Phi-3-small at 7B and Phi-3-medium at 14B for more demanding reasoning and math work, plus a multimodal Phi-3-vision variant for OCR and chart understanding.
Long context support, up to 128K tokens on some variants, and optimization through ONNX and DirectML made the family genuinely practical for local deployment rather than just a research curiosity, and it's available through Azure AI Studio, Hugging Face, and Ollama. Microsoft trained it on a carefully curated mix of web and synthetic data, with alignment and red-teaming work tied to its Responsible AI practices.
Phi-3 has since been superseded by Phi-4 and the 2026 Phi-4-reasoning-vision-15B release, both open source under the MIT license, which makes Phi-3 the previous generation in a lineup Microsoft has kept actively improving rather than a dead-end project.