Microsoft Finishes Pre-Training MAI-Base-2, a 2.8T-Parameter Model It Says Beats DeepSeek on Efficiency

According to a report by tech site tech-insider.org, Microsoft AI said on October 15 that it had finished pre-training MAI-Base-2, its in-house foundation model. Microsoft framed the run as a turning point.

MAI-Base-2 is a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active parameters per token. Microsoft says it was pre-trained natively with vision using an encoder-free vision-language architecture, meaning image understanding was built into the base model from the start rather than bolted on afterwards.

The training run is described in unusual detail: 16,000 Nvidia GB200 GPUs over 38 days across August and September 2026. Secondary reporting puts the corpus at roughly 50 trillion tokens, about 18 percent of it visual data, consistent with the native vision design.

The headline claim is not a benchmark score but efficiency. Microsoft says its recipe was 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base across four held-out evaluations, and 6.3 to 12.7 times more efficient than its own MAI-Base-1. The company also says MAI-Base-2 outperformed every open-weight model it tested, and that the compute it saved went into training a larger model.

The report adds that Microsoft AI CEO Mustafa Suleyman told staff at an internal town hall that early tests showed MAI-Base-2 beating the base model behind OpenAI's GPT-5.6 Sol on pattern-understanding and prediction tasks, while using roughly a quarter of OpenAI's pre-training compute.

None of that makes MAI-Base-2 a shipped product. The model is still in post-training, Microsoft has announced no public release date or pricing, and there is no finished chatbot built on top of it. Planned distribution is licensing to partners through Azure AI Foundry.

For scale, the report compares it with Meta's Muse Spark Open, an 816-billion-parameter MoE model released earlier in October. MAI-Base-2 is more than three times larger by total parameters, though with a lower share of active parameters per token, a trade-off aimed at keeping serving costs down.

The signal matters because Microsoft has leaned on OpenAI's models since 2019. Completing pre-training of an in-house foundation model is the clearest sign yet that it wants a credible alternative inside its own stack.

What to watch next: whether the efficiency numbers survive independent reproduction, when the model leaves post-training and becomes callable, and how it will be distributed through Azure AI Foundry.