\n<\/p>\n
Microsoft isn\u2019t like OpenAI, Google, and Meta; especially not when it comes to large language models<\/a>. While other tech giants prefer to launch multiple models almost overwhelming the users with choices; Microsoft launches a few, but those models always make it big among developers around the world. In their latest release, they have released 2 reasoning models: Phi-4-Reasoning and Phi-4-Reasoning-plus, both trained on the base Phi-4 model. The two Phi-4-Reasoning models compete with the mighty models like o1, o3-mini, and DeepSeek R1. In this blog, we will dive into the technical details, architecture, training methods, and performance of Phi-4-Reasoning models in detail.\u00a0<\/p>\n Let\u2019s explore the Phi-4-Reasoning models.\u00a0<\/p>\n Phi-4 is not new in the LLM world. This small and mighty language models broke the internet when it was launched last year. Now to cater to the increasing demand for reasoning models, Microsoft has released Phi-4-Reasoning models. These are 14 B parameters that excel at performing complex reasoning tasks involving mathematics, coding, and STEM questions. Unline its general purpose Phi-4 series, Phi-4-Reasoning is specifically optimized for long-chain reasoning \u2013 that is the ability to break down complex multi-step problems systematically into logical steps.<\/p>\n Also Read: Phi-4: Redefining Language Models with Synthetic Data<\/a><\/p>\n The two reasoning models released by Microsoft are:<\/p>\n The two 14B models currently support only text input, and Microsoft has released them as open-weight so developers can freely test and fine-tune them based on their needs. Here are some key highlights of the models:<\/p>\n For Phi-4 the team took several innovative steps involving data selection, its training methodology as well as its performance. Some of the key things they did were:<\/p>\n The Data Curation for training the Phi-4 reasoning models relied not just on sheer quantity but emphasized equally on the quality of data too. They specifically chose the data that was at the \u201cedge\u201d of the model\u2019s capabilities. This ensured that the training data was solvable but not easily.\u00a0\u00a0<\/p>\n The main steps involved in building the data set for Phi-4 models were:<\/p>\n For example, for the question \u201cWhat is the derivative of <\/strong>sin(<\/strong>x<\/strong>2 <\/strong>)<\/strong>2<\/strong>?\u201d; o3-mini gave the following output:<\/p>\n Step 2: Let u=x\u00b2 \u21d2 du\/dx=2x.\u00a0<\/p> Final Answer: cos(x\u00b2) * 2x.<\/p><\/pre>\n These artificially or synthetically generated chains of well-reasoned responses gave a clear blueprint on how a model should structure its own reasoning responses.\u00a0<\/p>\n The team essentially wanted the Phi-4-reasoning models to learn as they do, an approach that we humans usually rely on.\u00a0<\/p>\n Supervised Fine-Tuning (SFT) is the process of improving a pre-trained language model by training it on carefully selected input\u2013output pairs with high-quality responses. For the Phi-4-Reasoning models, this meant starting with the base Phi-4 model and then refining it using reasoning-focused tasks. Essentially, Phi-4-Reasoning was trained to learn and follow the step-by-step reasoning patterns seen in responses from o3-mini.<\/p>\n Using SFT during early training allowed the model to use Reinforcement learning<\/a> is teaching a model how to do better with feedback on all its generated outputs. The model gets a reward every time it answers correctly and is punished each time it responds incorrectly. RL was used to further train the Phi-4-Reasoning -Plus model. This training method refined the model\u2019s math-solving skills which evaluated the responses for accuracy and the structured approach.\u00a0<\/p>\nWhat is Phi-4 Reasoning?<\/h2>\n
Phi 4 Reasoning Models<\/h2>\n
\n
\n\n
\n \nDetails<\/th>\n Phi-4-Reasoning Models<\/th>\n<\/tr>\n<\/thead>\n \n Developer<\/strong><\/td>\n Microsoft Research<\/td>\n<\/tr>\n \n Model Variants<\/strong><\/td>\n Phi-4-Reasoning, Phi-4-Reasoning-Plus<\/td>\n<\/tr>\n \n Base Architecture<\/strong><\/td>\n Phi-4 (14B parameters), dense decoder-only Transformer<\/td>\n<\/tr>\n \n Training Method<\/strong><\/td>\n Supervised fine-tuning on chain-of-thought data; Plus variant includes additional Reinforcement Learning (RLHF)<\/td>\n<\/tr>\n \n Training Duration<\/strong><\/td>\n 2.5 days on 32\u00d7 H100-80G GPUs<\/td>\n<\/tr>\n \n Training Data<\/strong><\/td>\n 16B tokens total (~8.3B unique), from synthetic prompts and filtered public domain data<\/td>\n<\/tr>\n \n Training Period<\/strong><\/td>\n January \u2013 April 2025<\/td>\n<\/tr>\n \n Data Cutoff<\/strong><\/td>\n March 2025<\/td>\n<\/tr>\n \n Input Format<\/strong><\/td>\n Text input, optimized for chat-style prompts<\/td>\n<\/tr>\n \n Context Length<\/strong><\/td>\n 32,000 tokens<\/td>\n<\/tr>\n \n Output Format<\/strong><\/td>\n Two sections: reasoning chain-of-thought block followed by a summarization block<\/td>\n<\/tr>\n \n Release Date<\/strong><\/td>\n April 30, 2025<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n Key Features of Phi-4-Reasoning Models<\/h2>\n
Data Centric Training\u00a0<\/h3>\n
\n
Step 1: Apply the chain rule: d\/dx sin(u)=cos(u)*du\/dx.\u00a0\u00a0
\n
Supervised Fine-Tuning (SFT)<\/h3>\n
Training Details <\/h4>\n
\n
Reinforcement Learning<\/h3>\n
How does RL work?<\/h4>\n
\n