What it is
Phi is a family of small AI language models made by Microsoft. Where most famous models are enormous and live in giant data centres, Phi models are deliberately compact: small enough, in many cases, to run on an ordinary laptop, and in one case built into Windows itself. The popup on this table sums up the goal as strong performance on modest hardware.
"Small" here is relative. A model's size is usually counted in parameters, the learned numbers that hold everything it picked up in training. The biggest online assistants are thought to have hundreds of billions or more. Phi models have ranged from 1.3 billion to about 15 billion, with the popular "mini" versions at about 3.8 billion. That is small enough that a quantized copy, a compressed version explained on the parameters page, fits in a download of a few gigabytes.
The big idea: better lessons, not more of them. Most language models are trained by feeding them as much text from the internet as possible, much of it repetitive, low quality, or wrong. The Phi researchers bet on the opposite approach: pick the training material carefully, and write some of it on purpose, so that every example teaches something.
The idea has a charming origin. In an April 2024 feature, Microsoft described how its researcher Ronen Eldan was reading bedtime stories to his daughter and wondering how she had learned the words she knew. That line of thinking led to TinyStories, a May 2023 paper by Eldan and his colleague Yuanzhi Li. They had GPT-3.5 and GPT-4 write a large set of short stories using only words a typical three- or four-year-old understands, then trained very small models on them. Models with fewer than 10 million parameters, tiny by any standard, learned to write fluent, grammatical little stories. The lesson: a small model given a clean, focused curriculum can learn more than anyone expected.
The first Phi model, phi-1 in June 2023, applied that lesson to computer code, trained on "textbook quality" material, much of it written by another AI. That story, and the paper's memorable title, are told on the synthetic data page. Phi-2 and Phi-3 followed, and their benchmark results are covered on the parameters page. This page picks up the story from there.
The family at a glance. Microsoft releases Phi models in batches, often in several flavours at once. Dates below are from Microsoft's announcements or its model pages:
- Phi-3 (April 2024): phi-3-mini, about 3.8 billion parameters, followed by bigger siblings. Its paper reported running it fully offline on an iPhone 14 at more than 12 tokens per second.
- Phi-3.5 (August 2024): a mini model with better support for languages other than English, a vision model that can read images, and a mixture of experts model with 16 experts that uses 6.6 billion parameters at a time.
- Phi-4 (December 2024): 14 billion parameters, trained on 9.8 trillion tokens, with a strong focus on maths and reasoning.
- Phi-4-mini and Phi-4-multimodal (26 February 2025): a 3.8-billion text model and a 5.6-billion model that takes speech, images, and text together.
- Phi-4-reasoning, Phi-4-reasoning-plus and Phi-4-mini-reasoning (30 April 2025): versions trained to work through a problem step by step before answering, like the extended thinking models from bigger labs.
- Phi-4-mini-flash-reasoning (9 July 2025): a 3.8-billion reasoning model with a new internal design that Microsoft says gives up to 10 times higher throughput, aimed at phones and other small devices.
- Phi-4-reasoning-vision-15B (4 March 2026): a 15-billion model that reads images, charts, and screens, and can decide for itself whether a question needs step-by-step reasoning or a quick answer.
The line-up changes, so treat this list as correct at the time of writing and check Microsoft's Phi page for anything newer.
The licence. At the time of writing, Microsoft's Phi page says the Phi models are open source under the MIT licence, a short, permissive licence that lets you use, change, and share a model, including in a business, as long as you keep the licence notice. Two cautions. First, "open" here means open weights: you can download the finished model, but not the full training data or recipe. Second, it wasn't always this way. Phi-2 came under a non-commercial research licence and was relicensed as MIT on 6 January 2024. Phi-4 was first offered in December 2024 only on Microsoft's own cloud, for research, before its weights went up on Hugging Face under MIT on 8 January 2025. Always check the licence on the page of the exact model you download.
Phi Silica, the one built into Windows. At its Build conference on 21 May 2024, Microsoft announced Phi Silica, a Phi model made to run on the NPU, or neural processing unit, a chip designed for AI work, in its new Copilot+ PCs. VentureBeat reported it as a 3.3-billion-parameter model. Unlike the others, Phi Silica is not a download: it ships inside Windows, and app makers reach it through Windows' own programming tools, which Microsoft opened to developers from January 2025. Microsoft says it powers on-device features such as Click to Do and some rewrite and summarise features in Word and Outlook. Its documentation, updated in 2026, says Phi Silica can now also run on some ordinary PCs without an NPU, using the graphics chip instead, downloaded on demand. Availability varies by country and device, and the documentation notes it is not available in China.
Where you meet Phi. Most people meet it without knowing, through Windows features on Copilot+ PCs. Developers use it through Microsoft Foundry, Microsoft's cloud service for building with AI models (called Azure AI Foundry until Microsoft renamed it in November 2025), where Phi can be paid for per use. Anyone can download the open models free from Hugging Face or run them with Ollama, as in the steps below. At the time of writing, Microsoft's Phi page lists all three routes; check it for current options and prices.
Why it matters
It tests a big question about AI. For years the industry's working assumption was that bigger means better: more parameters, more data, more computing power. Phi is Microsoft's long-running experiment in the other direction, arguing that quality of data can make up for a lot of size. The Phi-4 paper describes a recipe "centrally focused on data quality", with about 50 kinds of synthetic dataset making up roughly 40 percent of the training mix. Whether the approach generalises beyond maths-style tests is exactly what critics question, as you'll see below. Either way, the argument has shaped how the whole field talks about training data.
Small models keep your data at home. A model running on your own laptop doesn't send your words anywhere. For a diary, a patient letter, a client contract, or an unfinished manuscript, that can matter more than having the cleverest model. Microsoft's Phi Silica documentation says prompts and responses stay on the device and are not sent to Microsoft during use. The privacy page covers the wider picture.
They work offline. No signal on a train, a flight, or a remote farm? A local model still answers. Microsoft's 2024 feature on Phi-3 gave the example of a farmer photographing a diseased crop and getting advice without a good network connection, an idea rather than a reported deployment.
They are cheap to run. A large model costs money for every request, whether you pay per token through an API or run expensive servers yourself. A small model on hardware you already own costs nothing per question. For businesses that make millions of simple requests, such as sorting emails or tagging support tickets, that difference adds up.
They use less energy. Every AI answer takes electricity, as the inference page explains. A smaller model does less arithmetic per word. Microsoft reported that on a Snapdragon X Elite laptop, Phi Silica's NPU processed a prompt using about 4.8 milliwatt-hours of energy. That's Microsoft's own measurement for its own chip, not an independent test, but the direction is not in doubt: small models on efficient chips use a fraction of the power of a big model in a data centre.
They are fast. With no round trip to a server, a small local model can start answering almost instantly, which matters for features like suggesting a rewrite as you type.
Real-world examples
Phi-4: from research preview to MIT licence (December 2024 to January 2025)
On 12 December 2024, Microsoft announced Phi-4, a 14-billion-parameter model. TechCrunch reported that at launch it was available only in limited form on Microsoft's cloud platform, for research purposes, and that Microsoft credited its maths strength to high-quality synthetic data plus carefully chosen human-written text. TechCrunch also noted it was the first Phi release since Sébastien Bubeck, a vice president who had been central to the Phi work, left Microsoft for OpenAI in October 2024.
The team knew critics would ask whether Phi had simply seen the test questions before. So in its technical report, Microsoft tested Phi-4 on the American Mathematics Competitions AMC-10 and AMC-12 held in November 2024, after all its training data had been collected, and reported that it outperformed larger models on those fresh problems. The same report was also frank about weaknesses. It said the model sometimes answers "Who is X?", where X is a plausible-sounding name, with an invented biography; that it is less good at following strict formatting instructions such as exact table layouts; and that it can over-explain simple problems.
Developers wanted to download it, not just try it on Microsoft's cloud. On 8 January 2025, Microsoft engineer Shital Shah announced that the official weights were on Hugging Face, the main website for sharing AI models, under the MIT licence, as VentureBeat reported. Phi-4's page there lists a 16,000-token context window and notes that the model was trained mostly on English and performs less well in other languages. The episode shows the usual Phi pattern: bold benchmark claims, an honest list of limits in the fine print, and eventually a permissive licence.
Phi Silica: a language model inside the operating system (May 2024 onwards)
When Microsoft announced Copilot+ PCs in May 2024, one piece got less attention than the headline features: a small Phi model, Phi Silica, would be built into Windows on every one of them, running on the NPU chip. This was a different way of meeting a language model. Nobody downloads Phi Silica or chats with it by name. It sits underneath features, and other apps can ask it for help.
In a December 2024 post, Microsoft said Phi Silica powered Click to Do (then in preview) and on-device rewrite and summarise features in Word and Outlook, and gave speed figures for Snapdragon X Elite laptops: about 230 milliseconds before the first word for short prompts, and up to 20 tokens per second. In its April 2025 "One year of Phi" post, Microsoft also mentioned Outlook summaries that work offline. Its 2026 documentation widened the reach to PCs without an NPU, running on the graphics chip, while warning that on those machines it will be slower, use more battery, and compete with games and video apps for the chip.
Phi Silica also shows the trade-off of built-in AI. It is private and free to use, but you don't choose the model, can't easily inspect it, and get the features Microsoft decides to ship. Microsoft's own notes warn that it can produce confident, wrong answers and should not be used as a sole source of facts.
An everyday scenario: the novelist on the train
This is an illustrative scenario, not a real case. Dana is revising a novel on a long train journey with patchy Wi-Fi. She has installed Ollama and Phi-4-mini on her laptop. She pastes in a clunky paragraph and asks for three tighter versions; the model does a decent job in seconds, with no connection and nothing leaving her machine. Then she asks it what year a real 1920s ocean liner sank, for a detail in her book. It answers confidently with a year. When she checks later, it's wrong. The small model was a good editor and a poor encyclopaedia, which is exactly the pattern the Phi researchers themselves describe.
Try it yourself
This exercise runs a Phi model on your own computer, for free, using Ollama, a program that downloads open models and runs them locally. It takes 20 to 40 minutes, mostly downloading. The goal is to see where a small model shines and where it guesses.
What you need, at the time of writing: a Windows, Mac, or Linux computer from the last few years, about 3 GB of free disk space, and ideally 8 GB of memory or more. Check Ollama's site for current requirements.
- Install Ollama. Go to ollama.com, click Download, choose your system, and install it like any other program. It is free to download and use at the time of writing.
- Open a terminal. On Windows, open PowerShell from the Start menu. On a Mac, open Terminal.
- Download and start Phi-4-mini. Type this and press Enter: > ollama run phi4-mini The first time, it downloads the model, about 2.5 GB at the time of writing. When you see a prompt, the model is running entirely on your computer.
- Give it a writing task. Paste this, adding your own text: > Here is a paragraph I wrote. Give me three tighter versions: one plain, one warm, one formal. Keep each under 60 words. [paste your paragraph] Small models are often good at this kind of rewriting, because everything they need is in front of them.
- Turn off your Wi-Fi and ask again. It still works. That's the privacy and offline advantage in action.
- Now ask a knowledge question you can check. For example: > Who won the Tour de France in 1987, and what nationality was he? (The answer is Stephen Roche, from Ireland.) Then try something more obscure that you know well, such as a fact about your own town or a minor historical figure. Note how confident it sounds whether it's right or wrong.
- Test the "Who is X?" weakness. Microsoft's own Phi-4 report says the model sometimes invents biographies for plausible names. Make up a name and ask: > Who is Margaret Ellsworth-Tane, the Scottish economist? If it writes a life story for someone who doesn't exist, you've just seen a hallucination.
- Try a tricky comparison. Ask:
Which is bigger, 9.9 or 9.11? Explain.The Phi-4 report mentions this kind of question as one its model can get wrong. Your result may differ. - Optional: try the larger Phi-4. If you have 16 GB of memory or more, type
/bye, thenollama run phi4. It is about 9.1 GB at the time of writing. Repeat steps 6 and 7 and compare. - Tidy up. Type
/byeto leave, andollama rm phi4-mini(andollama rm phi4) to free the disk space.
What to take away: use a small model for jobs where you supply the material, such as rewriting, summarising, or sorting your own text, and check any fact it produces from memory. For more on running models locally, see Parameters.
Common mistakes and misunderstandings
Assuming benchmark wins mean it beats big models at everything. Microsoft's announcements often compare Phi to much larger models on particular tests, mostly maths and reasoning. Those are Microsoft's results on Microsoft's chosen tests. In everyday use, the larger models usually know far more about the world.
Ignoring the overfitting question. In May 2024, Scale AI researchers wrote a fresh set of maths problems, GSM1k, and found that the Phi and Mistral families scored systematically higher on an older, public test than on the new one, a sign of overfitting, meaning partly learning the test rather than the skill. The same paper also found that Phi-2 still solved over half of the brand-new problems, so the models had learned real maths too. The story is on the benchmarks page. The fair conclusion is to treat any headline score as a hint and to test on your own tasks.
Expecting encyclopaedic knowledge. This is a trade-off the Phi team openly acknowledges. The Phi-3 paper said the model "simply does not have the capacity to store too much 'factual knowledge'", and suggested pairing it with a search engine. Training heavily on clean, textbook-style and synthetic material teaches reasoning well, but a small model has less room for trivia, names, and dates. Giving it the facts yourself, as in RAG or grounding, works much better than asking it to remember them.
Thinking synthetic data has no downsides. Text written by another AI can carry that AI's blind spots and style, and it covers what its authors thought to include. Critics argue this makes Phi models narrower than their scores suggest. Microsoft argues that careful curation outweighs the risk. Both points are worth holding; see synthetic data.
Assuming it works equally well in every language. Several Phi model pages say they were trained mainly on English and perform worse in other languages. Phi-3.5 and Phi-4-mini added more multilingual support, but test it in your language before relying on it.
Confusing Phi Silica with the downloadable Phi models. Phi Silica is built into Windows and isn't something you download from Hugging Face. The MIT licence applies to the Phi models Microsoft publishes for download, not to a component of Windows.
Thinking "MIT licence" means no responsibility. The licence gives broad permission, but your use still has to follow the law, and Microsoft's model pages warn against using Phi for high-stakes decisions such as medical, legal, or financial advice without extra safeguards. Features, models, and licence terms change, so check the current page for the model you use.
Assuming smaller always means worse. For a narrow, well-defined job, a small model that is fast, private, and free to run can be the better choice. Match the model to the task.
Related elements
- Gemma: Google's family of small open models, the closest comparison to Phi.
- Parameters: what "3.8 billion" means, the Phi-2 and Phi-3 results, and how to run models at home.
- Synthetic Data: the "textbook quality" training idea behind phi-1, told in full.
- Benchmarks: the GSM1k study that found signs of Phi overfitting a popular maths test.
- Inference: why running a model on your own device changes cost, speed, and energy use.
- Privacy: why keeping prompts on your own machine matters.
- Hallucination: the invented answers small models give when asked for facts they don't hold.
- Extended Thinking: the step-by-step reasoning approach behind Phi-4-reasoning.
- Llama: Meta's open family, another popular choice for running models locally.
- Qwen: Alibaba's open family, with several small sizes that compete with Phi.