phi-2

phi-2

A small text generation model that answers questions, continues chat dialogue and completes code. Made for researchers and developers who want a compact open model for safety studies and experiments.

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-2 is a small Transformer language model from Microsoft with about 2.7 billion parameters. It writes text from a prompt, in question and answer, chat or code layouts. It was trained on the same data as Phi-1.5 plus synthetic NLP texts and filtered websites. It has not been tuned with reinforcement learning from human feedback.

Who it suits: It suits researchers and developers who want a small, unrestricted model to study toxicity, bias and controllability, or to test ideas on modest hardware.

Industry | Who uses it | How they use it | Example
AI safety research | Research teams | Study toxicity, societal bias and controllability on a small open model | Probing how the model reacts to risky prompts
Education | Teachers and students | Run a small model locally to see how prompting works | Trying question, chat and code prompts side by side
Software | Developers | Prototype code completion or question answering | Completing a Python function from its comment

Skip it for production work, since the card says it was not tested for any production-level application.

Where to use it:

• Python scripts using the transformers library, version 4.37.0 or higher
• Windows, macOS and Linux
• A GPU, since the card's sample code sets cuda as the default device

Use case | Who | What they get back
Question answering | Researchers | A written answer after the question
Chat continuation | Developers | The next reply in a dialogue
Code completion | Developers | A function body after the comment

How to set it up:

• Python with a recent transformers library and torch
• About 5.6 GB of disk space for the F16 weights, or about 1.4 GB at 4-bit as an estimate
• A GPU for the card's sample code

1. Install the Hugging Face tools.
2. Download the model files, or let the sample code fetch them.
3. Run the sample code, or write your own prompt in one of the three formats.

pip install -U huggingface_hub
hf download microsoft/phi-2

How to use it: You write a prompt in the question, chat or code layout, and the model continues the text. For shorter answers, use the Instruct and Output layout. The card warns that an FP16 attention overflow can occur, and says older transformers versions need trust_remote_code set to True.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

torch.set_default_device("cuda")

model = AutoModelForCausalLM.from_pretrained("microsoft/phi-2", torch_dtype="auto", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2", trust_remote_code=True)

outputs = model.generate(**inputs, max_length=200)
text = tokenizer.batch_decode(outputs)[0]
print(text)

Format | What it does
Question | A standalone question, with the answer written after it
Instruct and Output | Asks for a more concise answer after the Output label
Chat | Names like Alice and Bob, and the model writes the next reply
Code | A function signature and comment, and the model writes the body

Example: You ask the model to compare mathematics to a lighthouse, using the Instruct line and an empty Output line. It continues after the Output label with a short written comparison. Treat the result as a starting point and check it yourself.

For newcomers and experts: Phi-2 is a small model that continues whatever text you start. A good first try is the card's lighthouse prompt, and then a short chat between two named people to see how it carries on.

If you know the field, this is a base-style Transformer without reinforcement learning from human feedback. That makes it a plain research subject for safety work. Expect the FP16 attention overflow the card describes, and expect its limits on accuracy for both code and facts.

Write a prompt in one of three formats
Tokenizer turns it into input
The model generates new tokens
Decode the tokens into text
Check the result yourself

Good to know:

• Accuracy: the model can produce wrong code and wrong statements, so treat output as a starting point.
• Production: the card puts direct production use without evaluation out of scope.
• Version: the card says transformers 4.37.0 or higher.
• License: MIT.
• Cost to run: none named.
• Data it reaches: it downloads model files from Hugging Face and works on the text you give it.
• Not for: production use without evaluation, or any reliance on its answers as definitive.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 2.8B
• Weights: about 5.6 GB in F16; about 1.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 182,126
• Likes: 3,536
• License: mit
• Languages: en
• Quantized versions: 64
• Fine-tunes: 416
• Tags: transformers, safetensors, text-generation, code

Links

• Model: https://huggingface.co/microsoft/phi-2
• Files: https://huggingface.co/microsoft/phi-2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download microsoft/phi-2

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