Microsoft Joins the Reasoning Race!!
AI Summary
Microsoft announced the release of 54 reasoning models, focusing on reasoning capabilities and the potential for small models trained on synthetic data. There are three models: 54 reasoning (14 billion parameters), 54 reasoning plus (enhanced with reinforcement learning), and 54 mini reasoning (3.8 billion parameters, derived from the DeepSeek R1 model). The models are primarily aimed at mathematical reasoning and have undergone various training phases, including distillation and alignment training, to improve their reasoning skills.
Key features of the models include:
- 54 reasoning: Base model with 14 billion parameters, trained on curated examples.
- 54 reasoning plus: Improved model using reinforcement learning.
- 54 mini reasoning: Distilled from a larger model, focused on providing decent performance with fewer parameters.
Microsoft is also exploring optimizing these models for local deployment on Windows devices, enhancing their efficiency and capabilities for applications such as Outlook.