Microsoft enters foundation model race
Key Questions
What foundation models has Microsoft announced?
Microsoft launched the MAI family including MAI-Thinking-1 (35B MoE), MAI-Image-2.5, and MAI-Code-1-Flash, marking a shift toward vertical integration beyond OpenAI reliance.
How do Microsoft's models perform on benchmarks?
MAI-Thinking-1 claims 97% on AIME and 53% on SWE-Bench Pro, beating Sonnet 4.6 in some areas, while image generation ranks #2 on leaderboards though independent verification is pending.
What tools support customization of Microsoft agents?
Frontier Tuning enables fine-tuning via RLEs for 10x efficiency matching GPT-5.4, alongside Scout persistent agents and policy conformance frameworks introduced at Build 2026.
Microsoft announces MAI and Mai-Flash foundation models, claiming to beat Claude and Gemini. This marks a strategic shift from relying on OpenAI to vertical integration. No independent benchmarks yet, but a major competitive signal in the model wars. Updated: Build 2026 reveals MAI-Thinking-1 reasoning model, Frontier Tuning for cost efficiency, and Scout agent. Image generation tops Google, but reasoning still trails DeepSeek. Now a concrete contender. New this run: MAI-Thinking-1 (35B MoE, beats Sonnet 4.6, 97% AIME, 53% SWE-Bench Pro), MAI-Image-2.5 (#2 leaderboard), MAI-Code-1-Flash (5B params, 51% SWE-Bench). Frontier Tuning enables customization. Mustafa Suleyman highlights transparency of MAI tech report. New this run: Mustafa Suleyman announces Frontier Tuning for custom agents (fine-tune via RLEs, matches GPT-5.4 with 10x efficiency).