AI Radar

Phi-4-reasoning

Available · Language, Reasoning

Phi-4-reasoning is a 14-billion-parameter open-weight reasoning model from Microsoft, created by supervised fine-tuning of Phi-4 and released on 30 April 2025 on Azure AI Foundry and Hugging Face under the MIT licence. Its answers contain a chain-of-thought section followed by a summary; the Phi-4-reasoning-plus variant adds a short reinforcement learning phase. [1] [2] [3] Primary source

Timeline of Phi-4-reasoning →

Claims and evidence

  • Released Primary source[1]page date and second paragraph [2]Model Summary: Release date
  • Status AvailableWeights remain downloadable from Microsoft's Hugging Face organisation; Microsoft Foundry lists Phi-4-reasoning as GA without a retirement date (checked 2026-10-01). Primary source[2]weights and licence (MIT) [4]Foundry Models from partners and community, Microsoft: Phi-4-reasoning, GA, no retirement date
  • Derived from (fine tune) Phi-4Supervised fine-tuning of Phi-4 on curated prompts and reasoning demonstrations. Primary source[1]Phi-4-reasoning and Phi-4-reasoning-plus [3]Abstract [2]Model Summary: Description, Architecture
  • Derived from (distillation) o3-mini (OpenAI) Primary source[1]supervised fine-tuning on reasoning demonstrations from o3-mini
  • Change · Reasoning Fine-tuned to write out a chain of thought before giving a summarized answer.Compared with Phi-4 Primary source[2]Model Summary: Outputs
  • Change · Context length Maximum length raised from 16K to 32K tokens to fit longer reasoning traces.Compared with Phi-4 Primary source[3]Increased Token Length
  • Change · Training data Supervised fine-tuning on about 1.4 million prompt-response pairs in math, coding and safety, with reasoning demonstrations generated by OpenAI o3-mini.Compared with Phi-4 Primary source[3] [1]Phi-4-reasoning and Phi-4-reasoning-plus
  • Input text Primary source[2]Model Summary: Inputs
  • Output textResponses consist of a chain-of-thought block followed by a summarization block. Primary source[2]Model Summary: Outputs
  • Open weights Yes Primary source[2]Model Summary: License MIT [1]Phi-4-reasoning and Phi-4-reasoning-plus
  • Context window 32k tokens Primary source[2]Model Summary: Context length [3]Increased Token Length
  • Parameters 14B Primary source[2]Model Summary: Architecture [1]Phi-4-reasoning and Phi-4-reasoning-plus
  • Access API, Open-weights download, On deviceAzure AI Foundry (api) and Hugging Face (open-weights-download) at release; ONNX-optimized versions for Snapdragon-powered Copilot+ PC NPUs followed on May 15, 2025 (on-device). Primary source[1]Phi-4-reasoning and Phi-4-reasoning-plus; Update: May 15, 2025

Lineage

Predecessors

No known predecessor.

Successors

No known successor.

Based on

  • Phi-4 · 12 December 2024 · derived (fine tune)
  • o3-mini · OpenAI · 31 January 2025 · derived (distillation)

Variants and derived

Siblings

None recorded.

All ancestors

All descendants

Variants

Phi-4-reasoning-plus

Same dates as Phi-4-reasoning.

  • Variant Inline variant in this record.Announced with Phi-4-reasoning. Microsoft trained it further with a short phase of outcome-based reinforcement learning; it generates on average about 1.5 times as many tokens as Phi-4-reasoning. Same 14B base and 32k context. Not listed in the Microsoft Foundry retirement schedule; weights on Hugging Face. Primary source[1]Phi-4-reasoning and Phi-4-reasoning-plus [3]Abstract [5]Model Summary

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Microsoft on AI Radar →