With Phi-4-Reasoning and its variants, Microsoft introduces more compact reasoning models, supposed to offer good performance on complex tasks while being less voracious in resources than large existing models.
Small, more energy -efficient language models are starting to attract attention more. Microsoft is part of this trend by applying techniques for reducing models to the elaboration of so -called reasoning, designed to solve complex problems by decomposing tasks into intermediate steps.
The firm thus presented Phi-4-Reasoning, Phi-4-Reasoning-Plus and Phi-4-Mini-Reasoning, a new range of compact language models, described as particularly efficient in mathematical reasoning. These models, according to Microsoft, gradually impose themselves as the backbone of so -called “agent” applications, capable of managing complex and multidimensional tasks. This capacity level has so far been reserved for very large models, still specifies the Redmond firm.
-The main model, Phi-4-Reasoning, is based on 14 billion parameters. Phi-4-Reasoning-Plus, which takes up its architecture while benefiting from additional training by learning by strengthening, making it possible to optimize the use of calculation resources during inference. According to Microsoft, these two models surpass the performance of competing models such as Openai O1-Mini and Deepseek-R1-Distill-Lalma-70B on several benchmarks, in particular mathematical reasoning and doctoral scientific questions.
Even less voracious, Phi-4-Mini-Reasoning, based on transforming type architecture, was specifically trained for mathematical reasoning step by step, based in particular on synthetic data generated by the Deepseek-R1 model.
All the models are available on Azure AI Foundry and Hugging Face platforms.
