AI news · Microsoft · models
Microsoft finishes MAI-Base-2, a 2.8 trillion parameter model
Microsoft says its own base model beats every open-weight rival it measured, and that it can license the weights to partners.
Microsoft is building large AI models of its own in earnest. In a blog post on Thursday 15 October 2026, Microsoft AI said it had finished pre-training MAI-Base-2. Pre-training is the long, expensive opening phase of building a model, in which it reads vast amounts of material and picks up how language and ideas fit together.
The model holds 2.8 trillion parameters, the numbers it adjusts as it learns. Only 80 billion of them work on each token, a word or piece of a word.
The claims are bold. Microsoft says it beats every open-weight model (one whose files anyone can download) that it ran through its own private question sets, kept out of training. It puts its pre-training efficiency at 4 to 12 times that of DeepSeek-V4-Pro-Base, the strongest open base model in its comparison.
Because none of its training data came from distilling other models (learning by copying their answers), Microsoft says it can offer the weights under licence to trusted partners through Foundry, its AI platform for businesses, Constellation Research reports.
At a staff meeting, Microsoft AI chief Mustafa Suleyman said the base model beats the one behind OpenAI’s GPT 5.6 Sol, Business Insider reports. He added that it took a quarter of OpenAI’s computing power to train.
Microsoft’s licensing plan points businesses on its cloud to a large in-house base model they can tune on their own data.
Sources
- Microsoft AI: Frontier Pre-training at Microsoft AI. Checked
- Constellation Research: Microsoft’s releases MAI-Base-2 aimed at open models with a twist. Checked
- Business Insider: Microsoft AI chief says the company beat OpenAI at its own game, using a quarter of the computing power. Checked
Alexandre S.
Alexandre S. is a creative ad strategist and copywriter with ten years' experience in digital marketing across multiple channels.
More AI news from
- India says lab AI agents read firms’ confidential files
- China drafts export curbs on 23 AI technologies
- Tim Cook says China has approved Apple Intelligence
- Claude cracks the 25-year riddle of NIST’s curve seeds
- Reflection releases Beam’s weights under an open licence
- Campaigns spend $137 million on AI ads, most with no label
- Anthropic says juniors are a shrinking share of its tech hires
- Trump’s Super Intelligence Force orders a report on AI risks
- Apple tightens App Store rules on AI nudify apps