Microsoft has announced its new in-house AI model, MAI-Base-2 (and I’m already really looking forward to the chance to test it)

The company Microsoft announced the completion of the pre-training stage for its new foundational artificial intelligence model— MAI-Base-2. The model was developed by Microsoft AI and is based on a mixture-of-experts architecture (Mixture of Experts, MoE), containing 2.8 trillion total parameters, of which the following are activated during inference: 80 billion.

Key specifications and training details:

• Scale and infrastructure: Training was conducted over 38 days in August and September 2026 on a cluster of 16,000 NVIDIA GB200 GPUs. The total dataset size was 50 trillion tokens.
• Out-of-the-box multimodality: В отличие от первой версии MAI-Base-1, модель изначально обучалась работе с текстом и изображениями с использованием архитектуры без отдельного визуального кодировщика. Визуальные данные составили 18% тренировочного корпуса (4,5 трлн токенов) и охватывают анализ графиков, пространственное мышление, чтение нот и медицинскую визуализацию.
• Compute efficiency: According to Microsoft's estimates, the new architecture outperforms the leading open-weight base model DeepSeek-V4-Pro-Base in FLOP efficiency by 4.2–12.3 times.
• Comparison with competitors: Microsoft AI conducted internal evaluations, and according to Microsoft AI CEO Mustafa Suleyman, MAI-Base-2 outperformed the base GPT-5.6 Sol model while using just a quarter of the compute.
• Data purity: The model was trained on an organic dataset without distillation (without using outputs from other AI models).

Will MAI-Base-2 be available in Microsoft Azure?

Yes, it will. And I'm really looking forward to testing it.

Thanks to the lack of distillation and full transparency regarding data provenance, Microsoft has officially announced its intention to license the model weights to enterprise customers and trusted partners through the Azure AI Foundry platform. In addition, the company plans to deploy future generations of products based on MAI-Base-2 Copilot and other Microsoft ecosystem services.

However, it is important to consider the project's current status:

• Current phase: The model has only completed pre-training and is now undergoing post-training, during which the base checkpoint is transformed into a production-ready chat product^[3].
• Timeline and pricing: At present, there are no publicly available API endpoints, specific Azure AI Foundry release dates, or pricing details to be announced.