Origin
Microsoft’s best-known open language models are the Phi series, small models that Wikipedia describes as open-weights and able to run on a user’s own device.
Microsoft was founded in 1975 by Bill Gates and Paul Allen and is headquartered in Redmond, Washington. Microsoft Research, created in 1991, employs more than 1,000 computer scientists and other researchers, according to Wikipedia. On Hugging Face the company lists speech, vision, code and language models side by side.
How it got here
Phi-1 came out on 21 June 2023 with 1.3 billion parameters. Phi-2 came next. Phi-3-mini, in April 2024, had 3.8 billion parameters, small enough to run on a phone. After Phi-3.5-mini came Phi-4, released on 12 December 2024 according to its model card, and Phi-4-mini, which its card dates to February 2025. In March 2026 Microsoft released Phi-4-reasoning-vision-15B, a vision-language model that decides when to reason before answering.
Phi is only part of the output. Microsoft has published the Florence-2 vision models; VibeVoice, a research model for generating long-form, multi-speaker speech from text; and a speech recognition model, VibeVoice-ASR. Older research models, such as the CodeBERT and WavLM encoders, are on the Hugging Face page too.
Licensing pattern
Most of what we track is MIT. The model cards for Phi-2, Phi-3-mini, Phi-3.5-mini, Phi-4 and Phi-4-mini list it, and so do the cards for Florence-2, VibeVoice and VibeVoice-ASR.
Some cards add use guidance on top of the licence. The VibeVoice card says the model is limited to research use and lists out-of-scope uses, including voice impersonation without recorded consent, disinformation and real-time voice conversion. It describes these as scenarios for which the release is not intended or licensed, and still names the MIT licence.
Not every older research model on the page states a licence in its card metadata. Check the individual repository before reuse.