LimitlessMindd/Phi-4-reasoning

TEXT GENERATIONPricing:Input $0.28 / Output $0.56Concurrent Unit Cost:1Model Size:14.7BQuant:FP8Context Size:32kPublished:Aug 21, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

Phi-4-reasoning is a 14.7 billion parameter dense decoder-only Transformer model developed by Microsoft Research, fine-tuned from Phi-4. It is specifically optimized for advanced reasoning, mathematical, scientific, and coding tasks, leveraging chain-of-thought traces and reinforcement learning. With a 32k token context length, it excels in memory/compute constrained environments and latency-bound scenarios requiring strong logical processing.

Loading preview...

Model Overview

Phi-4-reasoning is a 14.7 billion parameter language model developed by Microsoft Research, built upon the Phi-4 architecture. It is specifically fine-tuned for advanced reasoning capabilities across math, science, and coding domains. The model leverages supervised fine-tuning on chain-of-thought traces and reinforcement learning, using a blend of synthetic and high-quality filtered public data.

Key Capabilities & Features

  • Enhanced Reasoning: Designed to excel in complex problem-solving, particularly in mathematics, science, and coding.
  • Chain-of-Thought Output: Generates responses with a detailed reasoning process followed by a concise solution, aiding transparency and understanding.
  • Optimized for Efficiency: Suitable for environments with memory or compute constraints and scenarios requiring low latency.
  • Extensive Context Window: Supports a 32k token context length, allowing for deeper analysis and longer reasoning chains.
  • Robust Safety Alignment: Incorporates supervised fine-tuning with Microsoft's safety guidelines and undergoes extensive red-teaming.

Performance Highlights

Phi-4-reasoning demonstrates strong performance on various reasoning benchmarks, often outperforming larger open-weight models. For instance, it achieves competitive scores on AIME, GPQA-Diamond, OmniMath, and LiveCodeBench, showcasing its generalization to algorithmic problem-solving and planning tasks. It also shows significant improvements over its base model, Phi-4, on general abilities benchmarks like FlenQA, IFEval, and HumanEvalPlus.

Intended Use Cases

This model is primarily designed for research acceleration in language models and as a building block for generative AI features requiring strong reasoning and logic, especially in English. It is well-suited for applications where detailed thought processes and accurate solutions are critical.