Microsoft announces seven MAI models including MAI-Thinking-1 reasoning model with 35B active parameters and 256K context window

Microsoft presented seven new MAI models at Build spanning reasoning, code, image, speech transcription, and voice, led by MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5, and MAI-Voice-2. MAI-Thinking-1 uses no synthetic data or distillation from third-party models throughout its training pipeline.

MAI-Thinking-1 achieves 97% on AIME 2025 and 53% on SWE-Bench Pro, with blind human raters on Surge preferring it overall to Sonnet 4.6. MAI-Code-1-Flash achieves 51% on SWE-Bench Pro with 5B parameters. MAI-Image-2.5 reached #2 on Image Edit Arena with score 1401, +10 points over Nano Banana 2. MAI-Transcribe-1.5 achieves ~276x realtime speed with 2.4% AA-WER and ranks #3 overall on its leaderboard.

Microsoft released a 109-page technical report for MAI-Thinking-1 that drew praise from technically oriented readers for disclosing pipeline details, scaling ladder methodology, data curation, infrastructure metrics, and MFU numbers. One commenter called it 'one of the most transparent for a model at this scale.'

MAI-Transcribe-1.5 supports 43 languages and is priced at $6 per 1,000 minutes of audio via Microsoft Foundry. MAI-Code-1-Flash is available through GitHub Copilot and VS Code.