Phi-4-mini is a 3.8-billion-parameter open model from Microsoft, released in February 2025 under the MIT licence. It runs on an ordinary laptop (2.5 GB in Ollama), reads up to 128K tokens and handles summaries, classification, extraction and simple code. It knows fewer facts than big models, so give it the source text instead of asking from memory.
Other good fits
A first local model for learning: a small download, quick answers and a permissive MIT licence.
+ Pros
Pick it when the data must not leave the laptop: meeting notes, HR text, customer emails, draft reports.
Pick it for short, clear jobs: summarise, classify, pull fields out of text, reformat, or write a small Python script.
Pick it when hardware is modest. The 4-bit Ollama build is 2.5 GB and runs on a CPU; a GPU only makes it faster.
Pick it for longer documents than most small models accept: a 128K-token context in the default Ollama build.
− Cons
Pick something else when you ask for facts from memory. It is small and gets facts wrong; Microsoft suggests pairing it with search. AnythingLLM lets it answer from your own documents.
Pick something else when the task needs deeper reasoning or long agent runs: Qwen3.6-35B-A3B or gpt-oss-20b, if you have the memory.
Pick something else when you need to read images: Qwen3.6-35B-A3B takes image input.
Avoid if
You work mainly in a language other than English and quality matters. Microsoft says other languages perform worse.
Context
The reviewed model is microsoft/Phi-4-mini-instruct. Do not generalise this record to other Phi versions, quantizations or hosted endpoints.
Requirements
The card provides library and server examples. The checked primary sources do not provide a hardware, precision, context and workload-matched memory or latency measurement for this review.
Licence & cost
Prices checked 23 September 2026: free weights (MIT); you pay for the machine. A laptop with 8 GB of RAM runs the 2.5 GB 4-bit build; 16 GB leaves room for long inputs. The full 16-bit weights are 7.7 GB.
Microsoft also offers it as a hosted model on Azure, if you'd rather not run it yourself.
Running effort is low: one download, and the model doesn't change. Newer siblings exist for special jobs: Phi-4-mini-reasoning for maths, Phi-4-multimodal-instruct for images and audio.
The Ollama fp16 and q8_0 tags are set to a 4K context. Use the default tag for long documents.
First task
Documentation-based suggestion — not run here.
Read first task
Sort 50 customer emails by topic without sending them anywhere.
Save each email as a text file in a folder called emails.
In a terminal (macOS, Linux or Git Bash on Windows), run:
for f in emails/*.txt; do
label=$(ollama run phi4-mini "Label this email with exactly one of: billing, delivery, complaint, question, other. Reply with the label only. Email: $(cat "$f")")
echo "$f,$label"
done > labels.csv
Open labels.csv and check ten rows by hand. If more than one or two are wrong, add a one-line definition of each label to the prompt and run it again.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Based on Microsoft's Phi-4-mini-instruct model card, the Phi-4 collection on Hugging Face, the Phi-4-Mini technical report and Ollama's phi4-mini page, checked 23 September 2026. We did not run it for this review.
Phi-4-mini-instruct model card · Sep 23, 2026 — 3.8B parameters, 128K context, MIT licence, trained mainly on English with weaker other languages, limited factual knowledge (suggests search/RAG), vLLM and Transformers routes, Azure availability; first published 19 February 2025.
Microsoft Phi-4 collection · Sep 23, 2026 — Phi-4 family on Hugging Face, including Phi-4-mini-instruct, Phi-4-mini-reasoning and Phi-4-multimodal-instruct (replaces the old Phi-3 collection link).
Ollama: phi4-mini · Sep 23, 2026 — Default 3.8B build is 2.5 GB with a 128K context; fp16 (7.7 GB) and q8_0 tags use a 4K context.
Phi-4-Mini Technical Report · Sep 23, 2026 — Microsoft's technical report for Phi-4-mini and Phi-4-multimodal.
Microsoft Phi on Azure · Sep 23, 2026 — Microsoft's hosted offering of Phi models.
Further review notes & alternatives
Alternatives and trade-offs
gpt-oss-20b: the middle ground. Much stronger at reasoning, and OpenAI says it runs in about 16 GB of memory.
Qwen3.6-35B-A3B: far more capable, with image input, but needs about ten times the memory.
OpenVINO GenAI: the way to build Phi-4-mini into your own app on Intel laptops, using Intel's ready-converted version.
Run a compact local model for summaries, drafts, classification, and simple code.
Use its long context window to process sizeable documents locally.
Choose it for quick, scoped work on modest laptop hardware.
Microsoft · free · self-host
offline inference
cpu only
large context window
document summarization
code generation
text classification
mit licensed
tiny model
draft-docs
automate
open-source
write-code
talk-think
View sourced imageMicrosoft's published model-card benchmarks: four-row excerpt, typeset for readability. Publisher's internal evaluation; not independently reproduced and not a speed measurement. Checked 13 September 2026.
Microsoft Phi-4-mini-instruct model card · typeset excerpt
Owner-approved model-evidence illustration using the public official model card. Microsoft attribution retained. Playwright capture of an explicitly labelled typeset excerpt; source README SHA-256 c4f34916182a3ee87d95d00c92847cfa687ab21dc338a577275f0c34cf0e1dfd. Scores transcribed without alteration; only three compared models and four benchmarks selected from the full table.