---
title: Microsoft AI Unveils MAI-Base-2 Model with Significant Performance Improvements
url: https://www.elseif.net/microsoft-ai-unveils-mai-base-2-model-with-significant-performance-improvements
published: 2026-10-20T20:03:13+00:00
language: en
section: Models
source: https://www.blognone.com/node/151939
organizations: microsoft, mai-base-1, mai-base-2, mai-thinking-1, post-training, pre-training
publisher: elseif
---

# Microsoft AI Unveils MAI-Base-2 Model with Significant Performance Improvements

Microsoft AI has unveiled its new base model, MAI-Base-2, which shows significant performance improvements over its predecessor, MAI-Base-1. The MAI-Base-2 model is utilized during the pre-training phase before being fine-tuned into specialized models in the post-training phase. Previously, Microsoft introduced MAI-Thinking-1, a step-by-step reasoning model developed from MAI-Base-1, which had not been publicly released. This time, Microsoft has showcased MAI-Base-2 ahead of its progression to specialized models in the post-training phase.

MAI-Base-2 employs a Mixture-of-Experts (MoE) architecture and boasts a parameter size of 2.8 trillion, with an active count of 80 billion parameters. This represents a significant increase from MAI-Base-1, which had 1 trillion parameters and 35 billion active ones. Microsoft emphasizes the need to scale MAI-Base-2 to match the data volume, data quality, and computational resources required for training. The company has its own training data, negating the need to source results from other models. The training process utilized approximately 16,000 GPU instances over a span of 38 days.

The results show that MAI-Base-2 scores considerably higher on benchmark tests compared to MAI-Base-1. Furthermore, it outperforms open-weight base models from China, such as DeepSeek-V4-Pro-Base. Microsoft also compared MAI-Base-2 against other established open-weight models like Kimi K3 and Qwen3.8-2.4T-A95B, demonstrating superior performance even after those models underwent post-training.

Microsoft attributes its success to controlling the entire model pipeline from start to finish, coupled with its infrastructure, which enables the creation of high-quality models at lower costs. The models trained following MAI-Base-2 will be incorporated into other Microsoft products, and the company plans to offer them for rental to its customers.