Phi-4

tl;dr: Microsoft’s 14B open model. One size, 9.1GB on disk, 16K context. Built for reasoning on memory-constrained hardware — fits on a laptop, beats larger peers on math and logic benchmarks per Microsoft’s own evals.

What it is

phi4 is the Ollama tag for Microsoft’s Phi-4 release. One base size (14B parameters), text-only, 16K context. The README on ollama.com describes training data as a blend of synthetic datasets, filtered public-domain websites, and acquired academic books and Q&A datasets, with supervised fine-tuning and direct preference optimization on top.

The ollama.com page lists 14b as the single size, 7.7M cumulative pulls, updated roughly a year ago. It’s a single-purpose release: one chat-tuned 14B, no vision, no tools tag, no MoE. If you want a small reasoning model that fits on a workstation without quantization gymnastics, this is the shortlist.

Pull command

From the ollama.com detail page:

ollama run phi4

The default tag resolves to phi4:14b. To pull the same build explicitly: ollama pull phi4:14b.

Sizes and hardware

One size only. No quantization knobs surfaced on the library page beyond what Ollama applies by default.

License

The Ollama page does not surface a license link. Microsoft released Phi-4 upstream under the MIT license; verify at https://huggingface.co/microsoft/phi-4 before shipping commercial output. The Ollama README notes the model is “primarily in English” and lists developer responsibilities around accuracy, safety, and fairness — read those before high-risk deployments.

What the README says

The ollama.com README frames Phi-4 as a research accelerator and building block for generative AI features. Primary use cases Microsoft calls out:

• memory/compute-constrained environments
• latency-bound scenarios
• reasoning and logic

Out-of-scope use cases are listed as a general-purpose caution: developers are expected to evaluate accuracy, safety, and fairness for their downstream task. There’s no commercial restriction language on the Ollama page, but the upstream MIT terms still apply.

Gotchas

• 16K context only. Not a long-context model — Phi-4 was a pure reasoning release, not a context-length one.
• No tools tag, no function calling listed on the Ollama page. If you need tool use, look at the Llama or Qwen families instead.
• The Ollama page links to Hugging Face for weights but doesn’t quote a license on the page itself. Confirm upstream before production.
• Microsoft positions this for English-first workloads. Multilingual quality is not the headline.
• One size only — no 7B or smaller Phi-4 distillate published on Ollama as of this fetch.

Related: Llama 3.1 for the broader open LLM tier, DeepSeek-R1 for the open reasoning story at scale, Ollama for the runtime.

Untested here — facts from the ollama.com /library/phi4 page fetched 2026-09-10.