Microsoft AI, led by CEO Mustafa Suleyman, unveiled seven new in-house models — the "MAI family" — during the Build 2026 keynote on June 2, 2026, covering reasoning, code generation, image, voice and transcription, with all of them trained from scratch as a key selling point.
June 2, 2026 · Build 2026 Keynote · Microsoft AI
The MAI Family: Seven Models, Built From Scratch
Microsoft AI, led by CEO Mustafa Suleyman, unveiled an in-house lineup spanning reasoning, code, image, voice and transcription — trained on clean, commercially safe data and designed to work seamlessly together, reducing reliance on OpenAI and Anthropic.
7
models in the MAI family, built by a small team
~1T
total params in flagship MAI-Thinking-1 (35B active MoE)
128K
context length (up to 256K per some reports)
0
distillation from third-party models
MAI-Thinking-1 — benchmark scores
Microsoft's reported results across reasoning & code-fixing benchmarks
SWE-Bench Pro (code-fixing) ~52.8%
MAI-Code-1-Flash (5B) · SWE-Bench Pro 51%
On SWE-Bench Pro the flagship is reported on par with Claude Opus 4.6, and was preferred over Claude Sonnet 4.6 in blind human evaluations.
Efficiency on the MAIA 200 chip
MAI-Thinking-1 optimized for Microsoft silicon vs GB200
+30%
better performance per dollar
1.4×
better performance per watt
Praise
Efficiency, cost & deep Copilot integration
Logic of reducing OpenAI dependence
Code-1-Flash in VS Code, fast Transcribe
Frontier Tuning for proprietary models
Skepticism
Benchmarks compare vs older Opus/Sonnet
Scoring discrepancies on SWE-Bench Pro
Awaiting independent verification
The seven-model multimodal ecosystem
MAI-Thinking-1 · reasoningMAI-Code-1-Flash · codeMAI-Image-2.5 · imageMAI-Image-2.5-Flash MAI-Transcribe-1.5 · 43 langs, 5× fasterMAI-Voice-2 · 15 langsMAI-Voice-2-Flash
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…