Microsoft: Phi 4 Reasoning Plus

This 14B reasoning-focused model is built to solve math, science, and coding problems with clear, step-by-step solutions. Trained with supervised fine-tuning and reinforcement learning on specialized reasoning data, it delivers higher accuracy and transparent chains of thought. With a ~32K token context window, it can analyze technical documents and complex prompts, making it ideal for STEM Q&A, algorithm extraction, and tutorial-style assistance. Prompts that request “think step by step” improve reliability. Note that longer solution chains can increase latency, and performance is strongest in English. Use verification for code and logic outputs to ensure correctness in production workflows.