Tier 3 Practical open-weight
What the publisher claims: open-weight, open models, open-source
The one-liner: Practical small and mid-size models with unusually good artifact disclosure for a commercial lab.
The six questions
Every Phi-4 family checkpoint is downloadable from Hugging Face and GitHub, including the March 2026 phi-4-reasoning-vision-15B.
MIT across the family, with the model cards explicitly stating that no card content modifies the license, so use, modification, and redistribution are unrestricted including commercially.
Strong local deployment story: quantized Phi-4-mini runs in about 3GB of VRAM and Phi-4 14B on roughly a 12GB GPU, with official Ollama tags for every model.
The Phi-4 technical report documents the data composition and the synthetic-data curriculum in detail, model cards describe data sources and training hardware (for example 240 B200s over four days for the vision-reasoning model), and that release adds fine-tuning code and benchmark logs; the corpora themselves are not published.
Limited to moderate: an equivalent model can be approximated from the published methodology but the training data is not available.
Microsoft's Phi-4 family ships weights, fine-tuning code, and benchmark logs in public Hugging Face repos where the occasional outside PR is merged, but there is no public roadmap and Microsoft alone chooses the next release.
License
MIT across the Phi-4 family OSI-approved — license source ↗
- Model cards state that nothing in the card is a restriction or modification of the license, but the release is version-specific in practice
What is actually released
- weights (Phi-4 14B, phi-4-mini 3.8B, phi-4-multimodal 5.6B, phi-4-reasoning, phi-4-reasoning-plus, phi-4-reasoning-vision-15B)
- fine-tuning code for the reasoning-vision release
- benchmark logs
- model cards with data-overview and training-hardware detail
- technical report (arXiv 2511.19663) for the vision-reasoning model
Why it matters
Phi shows that smaller models can be more practically portable than frontier ones, and that a commercial lab can publish fine-tuning code and benchmark logs without opening the corpus.
Sources
- https://huggingface.co/microsoft/Phi-4-reasoning-vision-15B
- https://huggingface.co/microsoft/phi-4
- https://www.microsoft.com/en-us/research/blog/phi-4-reasoning-vision-and-the-lessons-of-training-a-multimodal-reasoning-model
- https://github.com/microsoft/Phi-4-reasoning-vision-15B
- https://www.microsoft.com/en-us/research/publication/phi-4-technical-report
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.