Phi-4

weights

open(MIT)

data

closed

code

closed

license

MIT(OSI)

Phi-4 ships MIT weights, genuinely permissive with no use restrictions. Neither the training corpus nor the training pipeline is published, however: Microsoft describes the synthetic data mix without releasing it, so the model is free to use but the recipe is not reproducible.

• https://huggingface.co/api/models/microsoft/phi-4 recorded 2026-09-29

  Hub API record with the license tag "license:mit", a license link to "https://huggingface.co/microsoft/phi-4/resolve/main/LICENSE" and six weight shards starting "model-00001-of-00006.safetensors"; the gated and private flags are both false.

• https://huggingface.co/microsoft/phi-4 recorded 2026-09-29

  Model card - "built upon a blend of synthetic datasets, data from filtered public domain websites, and acquired academic books and Q&A datasets"; the mixture is described, not released, and no training pipeline is linked.

• https://huggingface.co/microsoft/phi-4/raw/main/data_summary_card.md recorded 2026-09-29

  Data summary card - "Newly created synthetic, “textbook-like” data" alongside licensed sources; to "Have you concluded transactional commercial licensing agreement(s) with rights holder(s)" it answers Yes. No dataset download is linked.

• https://huggingface.co/microsoft/phi-4/raw/main/LICENSE recorded 2026-09-29

  The LICENSE file in the model repo is the MIT text - "Permission is hereby granted, free of charge, to any person obtaining a copy".

• https://huggingface.co/microsoft/phi-4/raw/main/README.md recorded 2026-09-29

  Raw model card - the only code link is the evaluation harness "https://github.com/openai/simple-evals"; searched the card for training code, pretraining scripts and a GitHub training repo: none.

• https://the-decoder.com/microsoft-releases-full-phi-4-model-with-weights-under-mit-license/ recorded 2026-09-29

  Report of the January 2025 release - "By releasing Phi-4 under the MIT license, Microsoft has given developers and researchers free rein to use, modify, and build upon the model".

• https://venturebeat.com/ai/microsoft-makes-powerful-phi-4-model-fully-open-source-on-hugging-face recorded 2026-06-04

  VentureBeat coverage of the Phi-4 MIT release; the host answers HTTP 429.

Adoption is measured as Hugging Face downloads for the single declared artifact, microsoft/phi-4.

• https://huggingface.co/api/models/microsoft/phi-4 recorded 2026-08-13

  626,702 downloads in the trailing 30 days for microsoft/phi-4

Phi-4 is a 14B model with outsized math and reasoning results, including 56.1% on graduate-level science questions and 80.4% on competition math, and its technical report says it surpasses GPT-4, its teacher, on STEM question answering. It remains a small, older model against the 2026 frontier and scores at the median of comparable models on Artificial Analysis's Intelligence Index.

• https://artificialanalysis.ai/models/phi-4 recorded 2026-09-29

  Phi-4 page - "Phi-4 scores 6 on the Artificial Analysis Intelligence Index, placing it below average among comparable models (median: 6)", ranked 45th of 75 on intelligence.

• https://arxiv.org/abs/2412.08905 recorded 2026-09-29

  Technical report abstract - "phi-4 substantially surpasses its teacher model on STEM-focused QA capabilities".

• https://huggingface.co/microsoft/phi-4/raw/main/README.md recorded 2026-09-29

  Model card benchmark table lists GPQA as "Complex, graduate-level science questions" and MATH as "Challenging competition math problems", with phi-4 at 56.1 and 80.4.

• https://the-decoder.com/microsoft-releases-full-phi-4-model-with-weights-under-mit-license/ recorded 2026-09-29

  Reports "achieving a 56.1 percent success rate on university-level questions and 80.4 percent on mathematical competition problems".