Description
Phi-4-reasoning-plus is an advanced reasoning model developed by Microsoft Research. It is fine-tuned from the Phi-4 model using supervised learning on a dataset of chain-of-thought traces and reinforcement learning. The model is designed to excel in reasoning tasks, particularly in math, science, and coding. It uses a dense decoder-only Transformer architecture with 14 billion parameters and is optimized for text inputs, especially in chat format. The model's training involved 16 billion tokens, with approximately 8.3 billion unique tokens, over a period of 2.5 days using 32 H100-80G GPUs. The model generates text responses with a reasoning chain-of-thought block followed by a summarization block, making it suitable for tasks requiring deep, multi-step reasoning or extensive context.
Read More
Phi-4-reasoning-plus has been evaluated on various benchmarks, including AIME, GPQA-Diamond, OmniMath, and LiveCodeBench, demonstrating strong performance in reasoning-intensive tasks. The model is also supported by frameworks like Ollama, llama.cpp, and any Phi-4 compatible framework. It is released under the MIT license and is intended for use in general-purpose AI systems and applications, particularly in memory/compute constrained environments and latency-bound scenarios.
The model incorporates a robust safety post-training approach via supervised fine-tuning, adhering to Microsoft's safety guidelines. It has undergone extensive safety evaluations, including adversarial conversation simulations and assessments by the independent AI Red Team at Microsoft. Despite its capabilities, developers should be aware of the model's limitations, such as potential biases, inaccuracies, and the elevated defect rate in election-critical queries. Developers are encouraged to apply responsible AI best practices and ensure compliance with relevant laws and regulations when deploying the model.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Phi-4-reasoning-plus Model Highlights
State-of-the-art reasoning model
Fine-tuned from Phi-4
14 billion parameters
Dense decoder-only Transformer architecture
Trained on 16 billion tokens
Supports chat format inputs
Generates reasoning chain-of-thought and summarization blocks
Robust safety post-training approach
Getting Started with Phi-4-reasoning-plus Model
Access page: Visit the Hugging Face model page
Load model: Download and load the Phi-4-reasoning-plus model
Configure environment: Set up the environment with required dependencies
Integrate: Use the model in your application
Fine-tune: Adjust the model for specific tasks if necessary
Phi-4-reasoning-plus Model's Use Cases
- Advanced reasoning
- Coding assistance
- Research acceleration
- General-purpose AI
- Memory-constrained environments
