Phi-4-reasoning

What it is

Phi-4-reasoning is a 14-billion-parameter, open-weight reasoning model released by Microsoft on 30 April 2025, alongside a technical report posted to arXiv^[1][2]. Open-weight means the trained parameters are published so that others can download and run the model themselves; this release came under the MIT licence^[1]. The model accepts up to 32K tokens of context^[1].

How it was made

It is a fine-tune of Phi-4, Microsoft's earlier small model. The company built it by supervised fine-tuning Phi-4 on a set of curated prompts that the report calls teachable, each paired with a reasoning demonstration generated by OpenAI's o3-mini^[2]. The model was therefore trained to imitate the step-by-step working of a stronger system, not just its final answers.

The fine-tuning data blended synthetic prompts with filtered public-domain web data focused on maths, science and coding, plus alignment data for safety^[1]. Training used 16 billion tokens, about 8.3 billion of them unique, and ran on 32 H100-80G GPUs for two and a half days between January and April 2025^[1].

A second variant, Phi-4-reasoning-plus, adds a short phase of outcome-based reinforcement learning^[2]. In reinforcement learning of this kind the model is rewarded for answers that come out correct, rather than for reproducing a stored reference solution.

How it answers

A response has two sections: a reasoning chain-of-thought block followed by a summarisation block^[1]. The plus variant produces longer reasoning traces, which the technical report links to its higher accuracy^[2].

Results

On AIME 2024, a US high-school maths competition, the base model scores 75.3 and the plus variant 81.3; on AIME 2025 the pair scores 62.9 and 78.0^[1]. On OmniMath, a maths benchmark, they score 76.6 and 81.9^[1]. On GPQA-Diamond, a set of graduate-level science questions, 65.8 and 68.9^[1]. On LiveCodeBench, a coding benchmark, 53.8 and 53.1 — the one measure where the plus variant is slightly behind^[1].

Microsoft said the two models match or beat much larger open models such as DeepSeek-R1-Distill-Llama-70B, and approach the full 671-billion-parameter DeepSeek-R1 on AIME 2025^[2].

Where it sits

Phi-4-reasoning belongs to the Phi family of small models and builds on Phi-4^[1]. Its language backbone was later used for Phi-4-reasoning-vision-15B in March 2026^[1]. The recipe it illustrates — a small base model, tightly filtered data, and reasoning traces distilled from a much larger system — is one route to reasoning behaviour without training a frontier-scale model from scratch.