{"rewrite":{"id":"r_91cdf2d95d3698dd9c01641a","clusterId":"c_dc2970e9853d2ec677d0a160","slug":"microsoft-announces-seven-ai-models-including-mai-thinking-1-and-mai-voice-2","model":"deepseek-v4-flash","headline":"Microsoft Announces Seven AI Models Including MAI-Thinking-1 and MAI-Voice-2","summary":"On June 2, 2026, Microsoft announced seven proprietary AI models, including the reasoning model MAI-Thinking-1, which the company says outperformed Anthropic's Claude Sonnet 4.6 in human evaluation, and the voice cloning model MAI-Voice-2. The models were developed using Microsoft's own Maia 200 chip and clean training data, with no distillation from other companies' models.","whyItMatters":"Microsoft is positioning its MAI model family as a self-sufficient alternative to leading models from Anthropic and OpenAI, claiming competitive performance without relying on third-party data or infrastructure.","webCardHtml":"\u003cp\u003eMicrosoft announced seven proprietary AI models on June 2, 2026, led by the reasoning model MAI-Thinking-1. The company claims the model outperformed Anthropic's Claude Sonnet 4.6 in human evaluation across 1,276 tasks, though benchmark scores show it trailing in several standard tests. MAI-Thinking-1 is a mixture-of-experts model with 1 trillion total parameters and 35 billion active parameters, trained on clean, licensed data and developed using Microsoft's own Maia 200 chip. The company stated it did not use distillation from other models.\u003c/p\u003e\u003cp\u003eThe lineup also includes MAI-Code-1-Flash, a 5-billion-parameter coding model that Microsoft says consistently beats Claude Haiku 4.5 on benchmarks; MAI-Image-2.5, an image generation model ranked third in text-to-image and second in image editing on the Arena leaderboard; a faster Flash variant; MAI Transcribe-1.5 supporting 43 languages; and MAI-Voice-2, a speech synthesis model supporting 15 languages. MAI-Thinking-1 is available in private preview on Microsoft Foundry, while MAI-Code-1-Flash will roll out to Visual Studio Code and GitHub Copilot.\u003c/p\u003e\u003cul\u003e\u003cli\u003e\u003cstrong\u003eMAI-Thinking-1\u003c/strong\u003e: MoE model, 1 trillion total parameters, 35 billion active; outperforms Claude Sonnet 4.6 in human evaluation\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eMAI-Code-1-Flash\u003c/strong\u003e: 5 billion parameters, high-speed coding, beats Claude Haiku 4.5 on benchmarks\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eMAI-Image-2.5\u003c/strong\u003e: Image generation, ranked 3rd in text-to-image, 2nd in image editing on Arena\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eMAI-Image-2.5 Flash\u003c/strong\u003e: Faster, more cost-efficient image generation variant\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eMAI Transcribe-1.5\u003c/strong\u003e: Transcription model supporting 43 languages\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eMAI-Voice-2\u003c/strong\u003e: Speech synthesis supporting 15 languages\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eMAI-Voice-2 Flash\u003c/strong\u003e: High-speed speech synthesis variant\u003c/li\u003e\u003c/ul\u003e","blueskyPost":"Microsoft dropped seven AI models yesterday, including MAI-Thinking-1, which it says beats Claude Sonnet 4.6 in human eval. Built on its own Maia 200 chip with clean data, no distillation. Also: MAI-Code-1-Flash, MAI-Image-2.5, and voice model MAI-Voice-2.","twitterPost":"Microsoft announced 7 AI models: MAI-Thinking-1 (claims to beat Claude Sonnet 4.6 in human eval), MAI-Code-1-Flash, MAI-Image-2.5, MAI-Voice-2, and more. Built on Maia 200 chip, clean data, no distillation.","threadsPost":null,"newsletterBlurb":"Microsoft announced seven proprietary AI models on June 2, including the reasoning model MAI-Thinking-1, which the company says outperformed Anthropic's Claude Sonnet 4.6 in human evaluation. The models were developed using Microsoft's own Maia 200 chip and clean training data, with no distillation from other companies' models. The lineup also includes a coding model, image generation models, transcription, and voice synthesis.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20260603-microsoft-ai-thinking/\",\"title\":\"Microsoft announces 7 AI models including 'MAI-Thinking-1' with performance comparable to Claude Sonnet 4.6 and voice cloning model 'MAI-Voice-2'\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4409,"outputTokens":1009,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1780458345,"createdAt":"2026-06-03T03:39:40.000Z","publishedAt":"2026-06-03T03:42:38.000Z","updatedAt":"2026-06-03T03:42:38.000Z"},"cluster":{"id":"c_dc2970e9853d2ec677d0a160","canonicalTitle":"MicrosoftがClaude Sonnet 4.6と同等性能な「MAI-Thinking-1」や音声クローンモデル「MAI-Voice-2」を含む7種のAIモデルを発表","representativeArticleId":"a_ba77021686b4fedf770fe635","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[],\"studios\":[],\"people\":[],\"type\":\"announcement\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-06-03T02:57:00.000Z","lastSeenAt":"2026-06-03T02:57:00.000Z","updatedAt":"2026-06-03T03:42:38.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20260603-microsoft-ai-thinking/","title":"MicrosoftがClaude Sonnet 4.6と同等性能な「MAI-Thinking-1」や音声クローンモデル「MAI-Voice-2」を含む7種のAIモデルを発表"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":[],"studios":[],"people":[],"type":"announcement","domain":"other","is_roundup":false},"keyFacts":["On June 2, 2026, Microsoft announced seven proprietary AI models, including the reasoning model MAI-Thinking-1 and the voice cloning model MAI-Voice-2.","Microsoft claims MAI-Thinking-1 outperformed Anthropic's Claude Sonnet 4.6 in human evaluation across 1,276 tasks, though benchmark scores show it trailing in several standard tests.","MAI-Thinking-1 is a mixture-of-experts model with 1 trillion total parameters and 35 billion active parameters, trained on clean, licensed data using Microsoft's own Maia 200 chip.","The lineup includes MAI-Code-1-Flash, a 5-billion-parameter coding model that Microsoft says consistently beats Claude Haiku 4.5 on benchmarks.","MAI-Thinking-1 is available in private preview on Microsoft Foundry, while MAI-Code-1-Flash will roll out to Visual Studio Code and GitHub Copilot."]}