One yr of Phi: Small language fashions making massive leaps in AI

Microsoft continues so as to add to the dialog by unveiling its latest fashions, Phi-4-reasoning, Phi-4-reasoning-plus, and Phi-4-mini-reasoning.

A brand new period of AI

One yr in the past, Microsoft launched small language fashions (SLMs) to clients with the discharge of Phi-3 on Azure AI Foundryleveraging analysis on SLMs to broaden the vary of environment friendly AI fashions and instruments out there to clients.

At this time, we’re excited to introduce Phi-4-reasoning, Phi-4-reasoning-plus, and Phi-4-mini-reasoning—marking a brand new period for small language fashions and as soon as once more redefining what is feasible with small and environment friendly AI.

Reasoning fashions, the following step ahead

Reasoning fashions are skilled to leverage inference-time scaling to carry out advanced duties that demand multi-step decomposition and inside reflection. They excel in mathematical reasoning and are rising because the spine of agentic functions with advanced, multi-faceted duties. Such capabilities are sometimes discovered solely in massive frontier fashions. Phi-reasoning fashions introduce a brand new class of small language fashions. Utilizing distillation, reinforcement studying, and high-quality knowledge, these fashions stability dimension and efficiency. They’re sufficiently small for low-latency environments but preserve sturdy reasoning capabilities that rival a lot greater fashions. This mix permits even resource-limited units to carry out advanced reasoning duties effectively.

Phi-4-reasoning and Phi-4-reasoning-plus

Phi-4-reasoning is a 14-billion parameter open-weight reasoning mannequin that rivals a lot bigger fashions on advanced reasoning duties. Skilled by way of supervised fine-tuning of Phi-4 on fastidiously curated reasoning demonstrations from OpenAI o3-mini, Phi-4-reasoning generates detailed reasoning chains that successfully leverage further inference-time compute. The mannequin demonstrates that meticulous knowledge curation and high-quality artificial datasets permit smaller fashions to compete with bigger counterparts.

Phi-4-reasoning-plus builds upon Phi-4-reasoning capabilities, additional skilled with reinforcement studying to make the most of extra inference-time compute, utilizing 1.5x extra tokens than Phi-4-reasoning, to ship larger accuracy.

Regardless of their considerably smaller dimension, each fashions obtain higher efficiency than OpenAI o1-mini and DeepSeek-R1-Distill-Llama-70B at most benchmarks, together with mathematical reasoning and Ph.D. stage science questions. They obtain efficiency higher than the total DeepSeek-R1 mannequin (with 671-billion parameters) on the AIME 2025 check, the 2025 qualifier for the USA Math Olympiad. Each fashions can be found on Azure AI Foundry and HuggingFace, right here and right here.

A graph of different colored bars — Determine 1. Phi-4-reasoning efficiency throughout consultant reasoning benchmarks spanning mathematical and scientific reasoning. We illustrate the efficiency beneficial properties from reasoning-focused post-training of Phi-4 by way of Phi-4-reasoning (SFT) and Phi-4-reasoning-plus (SFT+RL), alongside a consultant set of baselines from two mannequin households: open-weight fashions from DeepSeek together with DeepSeek R1 (671B Combination-of-Specialists) and its distilled dense variant DeepSeek-R1 Distill Llama 70B, and OpenAI’s proprietary frontier fashions o1-mini and o3-mini. Phi-4-reasoning and Phi-4-reasoning-plus persistently outperform the bottom mannequin Phi-4 by important margins, exceed DeepSeek-R1 Distill Llama 70B (5x bigger) and display aggressive efficiency towards considerably bigger fashions resembling Deepseek-R1.

A graph of numbers and a number of people — Determine 2. Accuracy of fashions throughout general-purpose benchmarks for: lengthy enter context QA (FlenQA), instruction following (IFEval), Coding (HumanEvalPlus), data & language understanding (MMLUPro), security detection (ToxiGen), and different normal expertise (ArenaHard and PhiBench).

Phi-4-reasoning fashions introduce a serious enchancment over Phi-4, surpass bigger fashions like DeepSeek-R1-Distill-70B and method Deep-Search-R1 throughout numerous reasoning and normal capabilities, together with math, coding, algorithmic drawback fixing, and planning. The technical report gives in depth quantitative proof of those enhancements by way of numerous reasoning duties.

Phi-4-mini-reasoning

Phi-4-mini-reasoning is designed to fulfill the demand for a compact reasoning mannequin. This transformer-based language mannequin is optimized for mathematical reasoning, offering high-quality, step-by-step drawback fixing in environments with constrained computing or latency. Advantageous-tuned with artificial knowledge generated by Deepseek-R1 mannequin, Phi-4-mini-reasoning balances effectivity with superior reasoning capability. It’s supreme for instructional functions, embedded tutoring, and light-weight deployment on edge or cell techniques, and is skilled on over a million numerous math issues spanning a number of ranges of problem from center faculty to Ph.D. stage. Check out the mannequin on Azure AI Foundry or HuggingFace in the present day.

A graph of numbers and a number of marks — Determine 3. The graph compares the efficiency of assorted fashions on widespread math benchmarks for lengthy sentence technology. Phi-4-mini-reasoning outperforms its base mannequin on lengthy sentence technology throughout every analysis, in addition to bigger fashions like OpenThinker-7B, Llama-3.2-3B-instruct, DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-Distill-Llama-8B, and Bespoke-Stratos-7B. Phi-4-mini-reasoning is similar to OpenAI o1-mini throughout math benchmarks, surpassing the mannequin’s efficiency throughout Math-500 and GPQA Diamond evaluations. As seen above, Phi-4-mini-reasoning with 3.8B parameters outperforms fashions of over twice its dimension.

For extra details about the mannequin, learn the technical report that gives further quantitative insights.

Phi’s evolution over the past yr has regularly pushed this envelope of high quality vs. dimension, increasing the household with new options to handle numerous wants. Throughout the dimensions of Home windows 11 units, these fashions can be found to run domestically on CPUs and GPUs.

As Home windows works in direction of creating a brand new kind of PC, Phi fashions have grow to be an integral a part of Copilot+ PCs with the NPU-optimized Phi Silica variant. This extremely environment friendly and OS-managed model of Phi is designed to be preloaded in reminiscence, and out there with blazing quick time to first token responses, and energy environment friendly token throughput so it may be concurrently invoked with different functions working in your PC.

It’s utilized in core experiences like Click on to Dooffering helpful textual content intelligence instruments for any content material in your display, and is accessible as developer APIs to be readily built-in into functions—already being utilized in a number of productiveness functions like Outlook, providing its Copilot abstract options offline. These small however mighty fashions have already been optimized and built-in for use throughout a number of functions throughout the breadth of our PC ecosystem. The Phi-4-reasoning and Phi-4-mini-reasoning fashions leverage the low-bit optimizations for Phi Silica and shall be out there to run quickly on Copilot+ PC NPUs.

Security and Microsoft’s method to accountable AI

At Microsoft, accountable AI is a basic precept guiding the event and deployment of AI techniques, together with our Phi fashions. Phi fashions are developed in accordance with Microsoft AI ideas: accountability, transparency, equity, reliability and security, privateness and safety, and inclusiveness.

The Phi household of fashions has adopted a strong security post-training method, leveraging a mixture of Supervised Advantageous-Tuning (SFT), Direct Desire Optimization (DPO), and Reinforcement Studying from Human Suggestions (RLHF) methods. These strategies make the most of numerous datasets, together with publicly out there datasets targeted on helpfulness and harmlessness, in addition to numerous safety-related questions and solutions. Whereas the Phi household of fashions is designed to carry out a variety of duties successfully, it is very important acknowledge that each one AI fashions could exhibit limitations. To higher perceive these limitations and the measures in place to handle them, please check with the mannequin playing cards under, which give detailed data on accountable AI practices and pointers.