Microsoft releases Phi-4-reasoning models

Background

The reasoning models released in April 2025 were built by fine-tuning Phi-4, a 14-billion-parameter model, rather than by training a new model from scratch.^[1][2]

What happened

On 30 April 2025 Microsoft released Phi-4-reasoning and Phi-4-reasoning-plus, both 14-billion-parameter reasoning models fine-tuned from Phi-4.^[1][2] Phi-4-reasoning was trained by supervised fine-tuning on reasoning traces from OpenAI's o3-mini. The plus version adds outcome-based reinforcement learning, which rewards correct final answers.^[1]

Key details

Phi-4-reasoning-plus scored 78.0 on AIME 2025 and 81.3 on AIME 2024, two sets of competition maths problems.^[2] The models have a 32K-token context and are released under the MIT licence.^[2]

Fine-tuning took 2.5 days on 32 H100 GPUs, using 16 billion tokens of training data.^[2]

Impact

Microsoft said both models outperform much larger open models, such as DeepSeek-R1-Distill-Llama-70B.^[1] The release showed that a comparatively small model, fine-tuned with reasoning examples and a final reinforcement stage, could be published as an open-weight reasoning model that developers could download and run.^[1][2]

Terms used here

A reasoning model is one trained to work through intermediate steps before giving a final answer. Fine-tuning means continuing the training of an existing model on a smaller, targeted dataset, so that it takes on new behaviour without being built from scratch. Reinforcement learning is a training method that adjusts a model according to the rewards its outputs earn; here, the reward is for reaching the correct final answer. AIME is a competition mathematics examination, and the 2024 and 2025 figures refer to two separate years of problems. DeepSeek-R1 is a larger open reasoning model against which the Phi-4-reasoning results were compared. Open weights means that the trained parameters are published for anyone to download, and the MIT licence governs how they may be used.

Competition mathematics problems are often used to score reasoning models because their answers can be checked automatically.