AI Radar

Phi-4-mini-reasoning

Available · Language, Reasoning

Phi-4-mini-reasoning is a 3.8-billion-parameter open-weight model from Microsoft for step-by-step math reasoning, released on 30 April 2025 on Azure AI Foundry and Hugging Face under the MIT licence. Built on Phi-4-mini with a 128K-token context, it is aimed at settings with limited compute or latency, such as education, embedded tutoring and edge or mobile devices. [1] [2] Primary source

Timeline of Phi-4-mini-reasoning →

Claims and evidence

  • Released Primary source[1]page date; Phi-4-mini-reasoning [2]Training, Model: Release date April 2025
  • Status AvailableWeights remain downloadable from Microsoft's Hugging Face organisation; Microsoft Foundry lists the model as GA without a retirement date (checked 2026-10-01). Primary source[2]weights and licence (MIT) [3]Foundry Models from partners and community, Microsoft: Phi-4-mini-reasoning, GA, no retirement date
  • Derived from (fine tune) Phi-4-miniMicrosoft applied a multi-stage recipe to Phi-4-Mini: mid-training on distilled long chain-of-thought data, supervised fine-tuning, rollout DPO and reinforcement learning with verifiable reward. Primary source[4]Abstract [2]Training, Model: Architecture
  • Derived from (distillation) DeepSeek-R1 (DeepSeek) Primary source[1]synthetic math data generated by DeepSeek-R1
  • Change · Reasoning Trained with distilled long chain-of-thought data, rollout DPO and reinforcement learning with verifiable reward to solve math problems step by step.Compared with Phi-4-mini Primary source[4]Abstract
  • Change · Languages Lists English as its only supported language, where the Phi-4-mini model card lists 23.Compared with Phi-4-mini Primary source[2] [5]Training, Model: Supported languages
  • Change · Other Designed and tested for math reasoning only, rather than for general-purpose use.Compared with Phi-4-mini Primary source[2]
  • Input text Primary source[2]Training, Model: Inputs
  • Output text Primary source[2]Training, Model: Outputs
  • Open weights Yes Primary source[2]licence: MIT
  • Context window 128K tokens Primary source[2]Model Summary; Training, Model: Context length
  • Parameters 3.8B Primary source[2]Training, Model: Architecture [1]Phi-4-mini-reasoning
  • 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-mini-reasoning; Update: May 15, 2025

Lineage

Predecessors

No known predecessor.

Successors

Based on

  • Phi-4-mini · 26 February 2025 · derived (fine tune)
  • DeepSeek-R1 · DeepSeek · 20 January 2025 · derived (distillation)

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

Variants

No variants recorded in this record.

Related AI Radar coverage

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

All model releases from Microsoft on AI Radar →