Phi-4

Phi-4

A 14B open language model from Microsoft Research, tuned for reasoning and chat in English with limited compute. It suits developers who add text features to apps or study mid-size models.

Made by microsoft on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Phi-4 is an open language model from Microsoft Research with 14 billion parameters, built as a dense decoder-only transformer. Its training mixed synthetic data, filtered public web pages, academic books and question sets, with a focus on reasoning. It was then refined with supervised fine-tuning and direct preference optimization so that it follows instructions closely.

Who it suits: It suits developers who add text features to products where memory, compute or response time are tight. Researchers who study language models can also use it as a building block.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Software | App developers | Add chat or reasoning to a product within limited hardware | Run it behind a support assistant that answers in English
Research | Language model researchers | Study a mid-size open model | Compare its behavior with other open models
Education | Learning tool builders | Build step-by-step help for math and logic questions | Return a worked explanation for a reasoning problem

If you need strong non-English output, you can skip it. The card says it targets English, and only about 8% of the training data is multilingual.

Where to use it: The record lists the transformers library, so Python is the main place. The card says it works best with prompts in chat format.

• Python with the transformers library
• Your own servers or workstation, with enough memory for the weights
• Windows, macOS or Linux with Python, to download the files

Title: Jobs it does
Use case | Who | What they get back
Memory-limited deployments | Developers | A capable text model that fits smaller setups than larger ones
Low-latency features | Product teams | Quick text replies in chat format
Reasoning and logic | Researchers | Written answers to reasoning tasks

How to set it up: You need Python with pip and the transformers library, plus memory for about 29.3 GB of weights in BF16. The visible part of the card gives no loading code, so follow the card for that.

• Python with pip
• The transformers library
• Memory for about 29.3 GB of BF16 weights, or about 7.3 GB at 4-bit by estimate

1. Install the Hugging Face download tool.
2. Download the model files to your computer.
3. Load the model through transformers, following its card.

pip install -U huggingface_hub
hf download microsoft/phi-4

How to use it: Send text prompts, preferably written as a chat with roles. The model returns generated text in reply. Its context length is 16K tokens, so keep prompts and answers within that.

Example: You send a chat message asking for a step-by-step explanation of why the sum of two even numbers is even. The model answers in text with the reasoning laid out in order. Check its answer before you rely on it.

For newcomers and experts: If you are new, this is a text model you run yourself. Start with a short chat prompt and read how it reasons. Then try a longer task.

If you know the field, the card says the training data extends Phi-3's mix with synthetic textbook-style data for math, coding and common sense. The model was trained on 9.8 trillion tokens across 1920 H100 GPUs. The trade-off is a small dense model chosen for quality data rather than size.

A chat prompt goes in
The dense transformer reads up to 16K tokens
It generates a text reply
You check the reply before using it

Good to know: It is a static model trained on an offline dataset, so its public data stops at its cutoff.

Title: Approximate weight size
BF16: 29.3 GB
4-bit estimate: 7.3 GB

• License: MIT
• Cost to run: none named
• Data it reaches: the model files from Hugging Face, and the prompts you send it
• Not for: high-risk uses without your own tests for accuracy, safety and fairness, and uses that break applicable laws

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 14.7B
• Weights: about 29.3 GB in BF16; about 7.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 236,972
• Likes: 2,343
• License: mit
• Languages: en
• Quantized versions: 173
• Fine-tunes: 127
• Tags: transformers, safetensors, phi3, text-generation, math, code, chat, conversational

Links

• Model: https://huggingface.co/microsoft/phi-4
• Files: https://huggingface.co/microsoft/phi-4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download microsoft/phi-4

Then follow the usage in its model card. Nothing in this kit runs by itself.