M

Microsoft Phi-4

Microsoft
16K
Context
$0.0600
Input /1M
$0.1400
Output /1M
N/A
Max Output
Run API Call Open in Chat

Microsoft Phi-4

Description

Phi-4 targets "complex reasoning tasks and operates efficiently with limited memory or when quick responses are needed." The 14-billion parameter model trained on synthetic datasets, curated websites, and academic materials. It emphasizes instruction-following accuracy and maintains safety standards, optimized for English inputs.

Technical Report: arXiv:2412.08905

Technical Specifications

Specification Value
Context Window 16,000 tokens
Parameters 14 billion
Context Length 16,384 tokens
Input Modalities Text
Output Modalities Text

Training Focus

  • High-quality synthetic data
  • Curated web content
  • Academic materials

Pricing

Type Cost (per million tokens)
Input $0.06
Output $0.14

Provider: NextBit (quantized as int4)

Capabilities

Phi-4 is designed for:

  • Complex reasoning tasks
  • Efficient operation with limited memory
  • Quick response generation
  • Instruction-following accuracy
  • Safe and responsible outputs

Optimized for: English language inputs

Supported Parameters

Parameter Supported
max_tokens Yes
temperature Yes
top_p Yes
stop Yes
frequency_penalty Yes
presence_penalty Yes
response_format Yes
Structured outputs Yes

Other Microsoft Phi models in the series:

  • Phi-3 series
  • Phi-2
  • Phi-1.5

Model Identity

Field Value
Name Microsoft: Phi 4
Model ID microsoft/phi-4
Short Name Phi 4
Author Microsoft Research
Created January 10, 2025
HuggingFace Slug microsoft/phi-4

Features

  • Tool Choice Support:
    • literal_none
    • literal_auto
    • literal_required
    • type_function
  • Structured Output Capabilities: Yes
  • Chat Completions: Enabled
  • Text Completions: Enabled

Access Points

Access Type URL
Chat Interface /chat?models=microsoft/phi-4
Model Comparison /compare/microsoft/phi-4
Model Weights HuggingFace (microsoft/phi-4)

Usage Statistics

Recent daily analytics show substantial adoption with millions of tokens processed across thousands of requests. Peak activity observed on December 5, 2025 and December 18, 2025.

Data scraped from LangMart on December 23, 2025