Phi models

Microsoft's small language models: Phi-4 (14B) and Phi-4-mini (3.8B) punch above their size on maths and reasoning, under the MIT licence.

Open weights
Models 6
Context up to 128K
Updated Nov 2026

Hugging Face Compare

On this page

At a glance
Models & versions
VRAM & download size
Run it locally
Best uses
Strengths & limits
About Phi
Guides
FAQ
Compare

At a glance

Type: Open-weight models
Developer: Microsoft (United States)
First release: Jun 2023
Models tracked: 6 (4 current)
Longest context: 128K tokens
Abilities: Tool calling
Licence: MIT licence, so commercial use is allowed.
Smallest local option: Phi-4-mini 3.8B · ~3.5 GB
Website: azure.microsoft.com
Hugging Face: huggingface.co/microsoft

Newest model

Phi-4-reasoning

Released Apr 2025 · 32K context

Smallest option

Phi-4-mini 3.8B

~3.5 GB VRAM · 2.5 GB download

Phi models and versions

Model	Released	Sizes	Context	Abilities	Licence	Status
Phi-4-reasoning Step-by-step reasoning (plus a "plus" variant)	Apr 2025	14B	32K	Text only	MIT	Current
Phi-4-mini-reasoning Maths reasoning on CPUs and small GPUs	Apr 2025	3.8B	128K	Text only	MIT	Current
Phi-4-mini Small model with tool calling and 128K context	Feb 2025	3.8B	128K	Tools	MIT	Current
Phi-4 14B model strong at maths and reasoning	Dec 2024	14B	16K	Text only	MIT	Current
Phi-3.5-mini Older small model	Aug 2024	3.8B	128K	Text only	MIT	Legacy
Phi-3 First Phi with a 128K context	Apr 2024	3.8B · 14B 3.8B, 14B	128K	Text only	MIT	Legacy

Sizes, context windows and abilities come from the Ollama library and Hugging Face model cards, checked 14 Nov 2026. Release months are the first public upload. Spotted a mistake? Tell us.

How much VRAM does Phi need?

Download sizes per Ollama build and the GPU memory each needs with an 8K context. On a CPU-only server the same number is the RAM you need. Longer contexts need more memory; work it out with the LLM VRAM calculator.

Model	4-bit (default)	8-bit	16-bit	VRAM (4-bit)	Fits on
Phi-4-reasoning 14B	11 GB	17 GB	29 GB	~13 GB	16 GB GPU
Phi-4-mini-reasoning 3.8B	3.2 GB	4.6 GB	7.7 GB	~4.5 GB	8 GB GPU
Phi-4-mini 3.8B	2.5 GB	4.1 GB	7.7 GB	~3.5 GB	8 GB GPU
Phi-4 14B	9.1 GB	16 GB	29 GB	~11 GB	12 GB GPU

VRAM = 4-bit download + about 10% runtime overhead + 0.5 GB for an 8K context, rounded up. Mixture-of-experts (MoE) models still need all their weights in memory, but they run much faster than a dense model of the same size, so they also work well split between GPU and system RAM.

Run Phi-4-reasoning locally

# Install Ollama (Linux), then pull and chat
curl -fsSL https://ollama.com/install.sh | sh
ollama run phi4-reasoning:14b

New to Ollama? Follow Install Ollama on Ubuntu 26.04 (or 24.04) for GPU drivers, the systemd service and remote access.

# Ollama serves an OpenAI-compatible API on port 11434
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "phi4-reasoning:14b", "messages": [{"role": "user", "content": "Hello"}]}'

# Full-precision weights from Hugging Face, OpenAI-compatible API on :8000
vllm serve microsoft/Phi-4-reasoning

vLLM loads the original 16-bit weights, so it needs roughly the 16-bit size above in GPU memory, plus room for the context. See Install vLLM on Ubuntu 26.04.

What Phi is good for

Maths and logic on small hardware

Phi-4-mini-reasoning runs on a CPU and solves multi-step maths.

Offline assistants

Phi-4-mini is small enough for laptops and supports function calling.

Strengths and limits

Strengths

MIT licence.
Very strong reasoning for the size, thanks to synthetic training data.

Watch out for

Phi-4 (14B) has only a 16K context.
Weaker general knowledge and multilingual skill than Qwen or Gemma.
No new Phi model on Ollama since 2025.

About Phi

Phi is Microsoft Research’s line of “small language models”. Starting with Phi-1 (2023), the team showed that carefully filtered and synthetic training data lets small models reason far above their size. Phi-3 (April 2024) brought a 3.8B model with a 128K context to phones and laptops.

Phi-4 (December 2024) is a 14B model aimed at maths and complex reasoning. In 2025 Microsoft added Phi-4-mini (3.8B, with function calling) and reasoning-tuned versions of both sizes. All Phi models are MIT-licensed.

Guides for running Phi

Tool LLM VRAM Calculator: How Much GPU Memory Do You Need? Updated Nov 2026 AI & LLMs How to Install Ollama on Ubuntu 26.04 and Run LLMs Locally Updated Nov 2026 AI & LLMs How to Install llama.cpp on Ubuntu 26.04 (CPU and NVIDIA GPU) Updated Nov 2026 AI & LLMs How to Install vLLM on Ubuntu 26.04 (OpenAI-Compatible LLM Server) Updated Nov 2026 AI & LLMs How to Install Open WebUI on Ubuntu 26.04 (Self-Hosted ChatGPT Alternative) Updated Nov 2026

Frequently asked questions

What is the latest Phi model?

The newest Phi model is Phi-4-reasoning, released in Apr 2025. It has a 32K-token context window.

How much VRAM do I need to run Phi locally?

The smallest option, Phi-4-mini 3.8B, is a 2.5 GB download at 4-bit and needs about 3.5 GB of GPU memory (or RAM on a CPU-only machine) with an 8K context. Larger sizes need more; see the hardware table on this page, and use the VRAM calculator for longer contexts.

How do I run Phi with Ollama?

Install Ollama, then run "ollama run phi4-mini:3.8b". Ollama downloads the 4-bit build and starts a chat; the same model is also available on the local API at http://localhost:11434.

Can I use Phi commercially?

MIT licence, so commercial use is allowed. Always read the licence of the exact model you deploy; it is linked in the models table.

Compare Phi

Phi vs Gemma

← All AI models