Microsoft Unveils 7 In-House MAI Models, Free of OpenAI
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Mustafa Suleyman told VentureBeat at Microsoft Build 2026 that Microsoft was "set free" from its OpenAI contract about six months ago to formally pursue superintelligence with its own researchers, data pipelines, and custom silicon
- Microsoft launched seven in-house "MAI" models, headlined by a 35-billion-active-parameter reasoning model (MAI-Thinking-1) trained from scratch on clean, commercially licensed data — explicitly without distillation from competitor systems
- The original 2019 Microsoft-OpenAI arrangement (cumulative investment exceeding $13 billion) had barred Microsoft from pursuing its own AGI research and capped model training size; a revised deal reported by Fortune and Axios in November 2025 removed those restrictions
- Microsoft's new "Frontier Tuning" capability lets enterprises customize MAI models with proprietary data in reinforcement-learning "training gyms," with an Excel-tuned MAI model reportedly matching GPT 5.4 performance at up to 10x greater efficiency
- Suleyman framed the pivot as strategic optionality rather than a rupture: "We have OpenAI, we have Anthropic, we have thousands of models inside Foundry," he said, with no immediate urgency to displace partners
- For the first time, MAI model weights are available for developer tuning through third-party platforms OpenRouter, Fireworks, and Baseten, signaling Microsoft is willing to distribute its AI beyond Azure-exclusive channels
Why it matters: With seven MAI models shipping and Frontier Tuning delivering up to 10x efficiency over GPT 5.4 for specialized enterprise tasks, Microsoft transforms from OpenAI's largest reseller into a direct frontier-model competitor — while maintaining the partnership. For enterprise buyers, the optionality now spans Microsoft's own stack alongside OpenAI and Anthropic within Foundry.
Ask SkimNews
