ArchiveStudio/Phi-4-mini-instruct

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.8BQuant:BF16Context Size:32kPublished:Jul 4, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

Phi-4-mini-instruct is a 3.8 billion parameter instruction-tuned decoder-only Transformer model developed by Microsoft, part of the Phi-4 family. Trained on 5 trillion tokens with a focus on high-quality, reasoning-dense synthetic data, it supports a 128K token context length. This model excels in reasoning tasks, particularly math and logic, and is optimized for memory/compute-constrained environments and latency-bound scenarios.

Loading preview...

Overview

Phi-4-mini-instruct is a 3.8 billion parameter instruction-tuned model from Microsoft's Phi-4 family, designed for efficiency and strong reasoning capabilities. It was built using synthetic data and filtered public websites, emphasizing high-quality, reasoning-dense content. The model incorporates supervised fine-tuning and direct preference optimization for precise instruction adherence and robust safety, and supports a 128K token context length.

Key Capabilities

  • Enhanced Reasoning: Demonstrates strong performance in math and logic, with notable scores on benchmarks like GSM8K (88.6) and MATH (64.0).
  • Multilingual Support: Features a larger vocabulary (200K tokens) and supports 23 languages, including Arabic, Chinese, French, German, Japanese, and Spanish.
  • Efficiency: Utilizes a new architecture with grouped-query attention and shared input/output embedding, making it suitable for memory/compute-constrained and latency-bound environments.
  • Instruction Following & Function Calling: Improved through advanced post-training techniques, supporting both chat and tool-enabled function-calling formats.
  • Competitive Performance: Achieves an overall score of 63.5 across popular aggregated benchmarks, outperforming several similar-sized models and even some 2x larger models in specific reasoning tasks.

Good For

  • General Purpose AI Systems: Suitable for broad commercial and research use.
  • Resource-Constrained Applications: Ideal for environments with limited memory or computational power.
  • Latency-Sensitive Scenarios: Designed for applications requiring quick response times.
  • Research & Development: Serves as a building block for generative AI features and accelerates research in language and multimodal models.
  • RAG Implementations: Can be augmented with search engines to address its inherent limitation in factual knowledge due to its compact size.

Finetunes
1 models