Microsoft Completes Pre-training of New MAI-Base-2 AI Model

Microsoft Refutes Claims of MAI-Base-2 Completion

Microsoft has not officially confirmed the development or completion of a model labeled “MAI-Base-2,” despite recent media reports suggesting the company has finalized its pre-training. While Crypto Briefing reported that the model’s pre-training is finished, the name does not appear on any official Microsoft model registry or verified product list.

Discrepancies in the MAI Model Series

Microsoft maintains a public-facing list of its AI models, and MAI-Base-2 is currently absent from these disclosures. The company has, however, confirmed its broader efforts in developing proprietary AI architecture. Microsoft AI recently introduced MAI-Thinking-1, an inference model that demonstrated high performance on the American Invitational Mathematics Examination (AIME), scoring 97.0% on the 2025 test and 94.5% on the 2026 version. This release confirms the existence of the “MAI” series, but it does not validate the existence of a “Base-2” iteration or its specific training status.

Microsoft continues to scale its proprietary computational infrastructure to support its model development pipeline. In an October 12 blog post, the company announced that its custom-designed accelerator, Maia 200, is already powering operational workloads for services like Copilot. Microsoft also noted that this hardware supports the internal MAI team’s efforts to build next-generation models. Furthermore, the company has scheduled the production deployment of the next-generation Maia 300 chip for 2027.

Lack of Technical Verification

Pre-training represents the initial phase of AI development where a model processes vast datasets to identify fundamental patterns. Because Microsoft has not released technical documentation, token counts, or specific hardware utilization data for a model named MAI-Base-2, the claims regarding its completion remain unverified by official sources. Reports regarding the model’s status lack the necessary technical disclosures—such as training resource allocation or specific architecture details—required to confirm the progress of such a large-scale project.

More on this story: Microsoft Launches Decision-1 AI Model with Free Output Tokens

More on Microsoft: Microsoft to Power AI Inference with New Maia 300 Chips from 2027 · Microsoft resets AI strategy and rebuilds Copilot · Microsoft Unveils Surface Laptop Ultra and Hybrid Intelligence Framework