Made by Microsoft · License: MIT
Model
Phi-4 (14B)
A free model for chat. You start it with one command.
Not sure which model? Start from your computer.
Start it
- Install Ollama from ollama.com/download.
- Open Terminal.
- Paste this command:
ollama run phi4:14bCopied word for word from its source.
What happens next
- It downloads about 9 GB the first time. Then type your message and press Enter.
- Type /bye to stop.
Will it run on my computer?
Mac
NVIDIA graphics card
AMD Ryzen AI Max
This is our estimate from the download size. A tested run on a computer overrides it. “Smaller version” means a smaller download than the command below. Tap your computer to see the exact command.
If it breaks
It forgets the start of a long chat, or ignores most of a document you pasted.
Ollama may not use the model's full context length by default. When a chat runs past that, the oldest messages are dropped without any warning.
Raise the context length with the slider in the Ollama app's Settings. Ollama's FAQ shows a chat command for this. Its example number, 4,096 tokens, is too small for long chats, so do not copy it. A longer context uses more memory. Check the CONTEXT column of
ollama psafter.Replies crawl out a word at a time.
The model did not fit in graphics memory, so Ollama quietly ran part of it on the CPU. Speed drops a lot, but you get no error.
Run
ollama pswhile the model runs. Look at the PROCESSOR column. If it does not say 100% GPU, free some memory. Close other models (Ollama's command list shows how), lower the context length, or pick a smaller version.ollama psCopied word for word from its source.
More help: Ollama's help page.
Sizes and versions
Will it run on my computer?
For phi4:14b (9 GB download; Same file as the counted RTX 4060 Ti 16 GB test, checked 2026-10-20). Context up to 16K, but Ollama may start lower, see below.
- Fit is our estimate from published download sizes. A tested run overrides it.
- Fits = download x 1.25 + 2 GB fits in usable memory. That leaves room for longer chats.
- Tight = only download + 1.5 GB fits. Short chats work, but long chats or documents may spill onto the CPU and slow down.
- Usable memory on a graphics card is all of its graphics memory.
- On a Mac it is 2/3 of memory up to 32 GB and 3/4 at 64 GB+. By default macOS keeps part of memory away from the GPU. Apple says a 32 GB Mac gives the GPU 21 GB and a 64 GB Mac 48 GB.
- On an AMD Ryzen AI Max+ PC with 128 GB, it is the graphics memory you set in AMD Software. AMD says you can set up to 96 GB.
- Apple · AMD
Sizes checked 2026-10-20.
START HERE Easy: Ollama
ollama run phi4:14bCopied word for word from its source.
Recommended quants
- Ollama phi4:14b (9.1 GB)
ollama phi4Ollama default tag.
Scores and how it compares
Independent score: Artificial Analysis Intelligence Index 6* (estimate), checked 2026-10-20 · source.
Artificial Analysis marks this score as an estimate, so it does not rank.
Arena (ranks our computer pages):
- Arena Text 1256 ±5 (Phi-4)
- Arena Writing 1210 ±10 (Phi-4)
Checked 2026-10-19.
Pick this if you have a 16 GB graphics card and want a model for chat or writing.
Skip this if you need coding, documents or images help. It has no counted score for those jobs on our pages. Its context window is 16K tokens.
Compare: Gemma 4 (26B-A4B / 31B) · Pick your computer
Settings and apps
Chat template Works
Ollama's tag carries the chat template.
Sampling settings
Runtimes
| Ollama | Works | Official library tag phi4:14b. |
|---|
Sources
A 14B model from Microsoft. One Ollama command (about 9 GB). It fits a 16 GB graphics card.
Known issues
Sources checked
Spotted something wrong or a new fix? Email corrections to hello@runcards.dev. Corrections are reviewed before publishing.
Details
Words used here
- Arena (LMArena) is a public leaderboard. People compare two answers and vote for the better one. WebDev, Coding and Writing are some of its boards.
- B after a number means billion weights (parameters). A model with more weights needs more memory. T means trillion.
- Chat template (often a Jinja template) tells the program how to lay out the chat for the model.
- Context (also called the window) is how much text the model can keep in mind at one time. It is counted in tokens.
- CPU is the main chip of the computer. A model that runs on the CPU is much slower than on the GPU.
- GPU is the graphics chip. AI models run fastest on it. On a Mac, the GPU is part of the main chip.
- K after a token count means thousand, and M means million. 32K is about 32,000 tokens.
- MoE (mixture of experts) is a model that uses only part of its weights for each token, so it is fast for its size. A name such as 26B-A4B means 26 billion weights in total, with 4 billion active for each token. A dense model uses all its weights for each token.
- Quant (quantized version) is a smaller copy of a model. It keeps each number in fewer bits. It uses less memory and can lose a little quality.
- Tag is the part after the colon in a model name, such as gemma4:26b. It picks one size or version.
Checked Oct 20, 2026. Some staleness is always possible: versions and links change. RunCards is not affiliated with any model maker, recipe author or repo. Troubleshooting belongs to each repo or author. Read more
Did you find a mistake? Suggest a fix. See what we changed.