Phi-4-mini-instruct
MicrosoftPhi-4-mini-instruct is a 3.8B-parameter dense decoder-only language model with a 128K-token context window, tuned for chat and instruction following through supervised fine-tuning and direct preference optimization. Microsoft positions it for memory-constrained and latency-bound settings. It belongs to the Phi-4 family, alongside reasoning and multimodal variants.
Openness
3 high confidence- weights
- open(MIT, on HF)
- data
- closed(synthetic-heavy training data undisclosed)
- code
- closed(no training pipeline)
- license
- MIT(OSI permissive)
MIT-licensed open weights, with the training and post-training data and the training code unreleased, so this is open weights rather than open source.
- https://huggingface.co/api/models/microsoft/Phi-4-mini-instruct recorded 2026-09-29
Repo tags include "license:mit"; the file list has "model-00001-of-00002.safetensors" and no training pipeline. Searched the file list for train and pipeline code: only sample_finetune.py.
- https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/main/LICENSE recorded 2026-09-29
LICENSE file is the MIT License: "Permission is hereby granted, free of charge, to any person obtaining a copy".
- https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/main/README.md recorded 2026-09-29
Model card describes the data mixture in prose: "training data includes a wide variety of sources, totaling 5 trillion tokens"; it links no corpus or training pipeline. Searched the body for released dataset and training code: none.
Adoption
3 high confidenceAdoption is read from Hugging Face downloads on the one declared artifact, microsoft/Phi-4-mini-instruct.
- https://huggingface.co/api/models/microsoft/Phi-4-mini-instruct recorded 2026-08-13
429,490 downloads in the trailing 30 days for microsoft/Phi-4-mini-instruct
Capability
2 high confidencePhi-4-mini-instruct is strong on math for a 3.8B model (GSM8K 88.6) but weak on graduate-level science questions (GPQA 25.2), and it trails GPT-4o-mini overall on the model card's quality table. It sits well below current frontier chat models.
- https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/main/README.md recorded 2026-09-29
Quality table, Phi-4 mini-Ins column: "MMLU-Pro (0-shot, CoT)" 52.8, "GPQA (0-shot, CoT)" 25.2, Overall 63.5 against 75.5 for GPT-4o-mini; the card says "the model with only 3.8B-param achieves a similar level of multilingual language understanding" as larger models.
Verified 2026-09-27