Skip to content
← Models
Open weightsMicrosoft/10 trustVerified
Phi-4-reasoning logo

Phi-4-reasoning

Microsoft's 14B MIT-licensed reasoning model that punches above its size on math and science.

Visit official page →

About

Phi-4-reasoning is a 14B-parameter open-weight reasoning model from Microsoft Research, released April 30, 2025, fine-tuned from Phi-4 with chain-of-thought SFT and reinforcement learning. Its published scores beat the 70B DeepSeek-R1 distill on AIME 2024/2025 and match or approach much larger models on OmniMath and GPQA-Diamond. It ships under the permissive MIT license and runs in Ollama, vLLM and llama.cpp; honest limits are a modest 32K context, English-first training, and outputs designed around a specific think-then-answer prompt format.

Specs

Context window

32K tokens

Pricing

Free, self-hosted

Free quota

HuggingFace free inference; fully self-hostable at no cost

API access

Hugging Face, Ollama, vLLM, llama.cpp, Azure AI

Weights

Open weights

Released

April 30, 2025

Benchmarks

Microsoft HF model card (provider-published):

  • AIME 2024: 75.3 | AIME 2025: 62.9
  • OmniMath: 76.6
  • GPQA-Diamond: 65.8
  • LiveCodeBench: 53.8
  • HumanEvalPlus: 92.9 | MMLUPro: 74.3 | IFEval Strict: 83.4 | ArenaHard: 73.3

Outperforms DeepSeek-R1-Distill-70B (AIME 2024 69.3, AIME 2025 51.5) and o1-mini (AIME 2024 63.6, GPQA-D 60.0) on most reasoning benchmarks.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

coderesearchdataothers

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier, metrics-v1). Evidence: Full HF model card read: MIT license, 14B, 32K context, release date 2025-04-30 confirmed; Provider-published benchmark table cross-checked against VentureBeat reporting (scores match); MIT license = full commercial use, fine-tuning and distillation allowed; Follow-up Phi-4-reasoning-plus (RL, higher accuracy) released alongside; works out of the box in Ollama/vLLM. License: MIT License (permissive; commercial use, modification, and distillation allowed)..

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
7.0/10
Maintenance · recency, releases, issue load
7.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
9.0/10
  • · Full HF model card read: MIT license, 14B, 32K context, release date 2025-04-30 confirmed
  • · Provider-published benchmark table cross-checked against VentureBeat reporting (scores match)
  • · MIT license = full commercial use, fine-tuning and distillation allowed
  • · Follow-up Phi-4-reasoning-plus (RL, higher accuracy) released alongside; works out of the box in Ollama/vLLM

Sources