Microsoft finishes pre-training MAI-Base-2, its next in-house foundation model

The successor to MAI-Base-1 has cleared its most compute-heavy phase as Microsoft builds out its own AI stack

Microsoft has finished pre-training MAI -Base-2, the next foundation model from its in-house AI group.

Pre-training is the most compute-intensive stretch of building a large language model. It produces the raw intelligence, not the polished product, and it signals that Microsoft’s homegrown AI effort is moving into its second generation.

Think of pre-training as a model’s undergraduate degree. It reads a vast library of text and learns to predict what comes next, picking up grammar, facts and patterns along the way.

What it can’t do yet is follow instructions gracefully or reason through a problem on request. Those skills come in later phases, typically called mid-training and post-training, where the raw model gets shaped into something people can actually use.

The name gives a hint about where it fits. Its predecessor, MAI-Base-1, is the base model underneath MAI-Thinking-1, the flagship reasoning model Microsoft AI introduced earlier this year.

MAI-Base-1 set a high bar on scale. It was pre-trained on 30 trillion tokens of clean, human-generated data, then went through another 3.55 trillion tokens during mid-training.

The hardware was equally serious. Microsoft trained that first base model on 8,192 NVIDIA GB200 GPUs, a cluster built for exactly this kind of marathon workload.

The MAI program went public in a big way on June 2, 2026, at Microsoft Build. Microsoft AI, the in-house team led by Mustafa Suleyman, rolled out a lineup of seven models at once.

Daily. Free. Join 34,000+ readers across crypto, finance, and policy.

The headline claim was independence. Microsoft said every model, including MAI-Thinking-1, was built from the ground up without using third-party models for distillation.

The lineup went beyond text. Microsoft’s MAI family covers multimodal work, including transcription, voice and image processing.

On performance, MAI-Thinking-1 has posted around 52.8% on SWE-Bench Pro, a benchmark built on real-world software engineering tasks. It also scored 97% on AIME 2025, a test drawn from a demanding math competition.

Microsoft is running MAI models across its own workloads, including Copilot, its AI assistant woven through products like Windows and Office.

The development approach is described as a hill-climbing training pipeline, where each training cycle adjusts data quality and scale, and the improvements carry forward.

Hardware is one piece to watch. MAI-Base-1 was trained on NVIDIA GB200s, but Microsoft is expanding infrastructure for its own Maia 200 and 300 chips. The company plans to move toward multi-gigawatt Maia deployments starting in 2027.

There are a few concrete things to track from here. First, whether Microsoft discloses the data and compute behind MAI-Base-2, and how those numbers compare with the 30 trillion tokens and 8,192 GPUs used for its predecessor.

Second, whether a reasoning model built on MAI-Base-2 can move beyond the around 52.8% that MAI-Thinking-1 posted on SWE-Bench Pro.

Third, timing. If MAI-Base-2’s descendants land around the same window as the planned 2027 Maia expansion, Microsoft would have a new generation of models arriving alongside a new generation of its own hardware.