The Extended Brief

Microsoft's releases MAI-Base-2 aimed at open models with a twist

Brief by The AI News AI newsroom · Oct 16, 2026, 7:02 AM EDT edition

Original reporting by Constellation Research — Larry Dignan · published Oct 16, 2026, 6:45 AM EDT

Enterprises wanting to train on their own data may get a new licensable base model from Microsoft — but despite the open-model comparisons, MAI-Base-2's weights are licensed, not open.

Key points

• Microsoft released MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters. source ↗
• Microsoft said it will license MAI-Base-2's weights, a potential route for enterprises to train models on their own data. source ↗
• Microsoft says MAI-Base-2 is more efficient than Chinese open-weight rivals and was trained without distillation. source ↗
• Microsoft's documentation compares MAI-Base-2 to open-weight models including Alibaba's Qwen, DeepSeek, and Kimi. source ↗
• Microsoft AI CEO Mustafa Suleyman argued the open-versus-closed case on X, where some called the comparison a joke. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• If your team fine-tunes open-weight models like Qwen or DeepSeek, request MAI-Base-2's license terms to compare total cost against your current training stack.
• Ask Microsoft for evidence behind its 'strong data provenance' claim on training tokens before routing proprietary data through its pipeline.
• Benchmark MAI-Base-2 against Qwen, DeepSeek, and Kimi on your own workloads rather than relying on Microsoft's in-house comparisons.

Context

Mixture-of-experts models route each token to a small subset of parameters, so MAI-Base-2 activates 80 billion of its 2.8 trillion parameters at a time to cut compute cost. Open-weight models such as Alibaba's Qwen and DeepSeek publish their parameters for anyone to download and fine-tune; Microsoft is instead licensing access. Distillation, a common shortcut where a smaller model learns from a larger one's outputs, is something Microsoft says it did not use.

What to watch

• The actual license terms — price, redistribution rights, fine-tuning scope — will determine whether enterprises treat MAI-Base-2 as a real open-weight alternative.
• Independent benchmarks against Qwen, DeepSeek, and Kimi will confirm or undercut Microsoft's efficiency claims.

Related briefs

• ASML Is Building a Hybrid Bonding Tool!
• Open weight takes 71.5% of tokens and 23% of spend
• Accenture Tells Employees They Can Now Spend Only $29 on Claude Per Month
• Character.ai was worth US$2.7B to Google. Disney just offered US$20M

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.