惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

D
DataBreaches.Net
罗磊的独立博客
雷峰网
雷峰网
量子位
V
Visual Studio Blog
Vercel News
Vercel News
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
The Cloudflare Blog
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
宝玉的分享
宝玉的分享
月光博客
月光博客
Martin Fowler
Martin Fowler
aimingoo的专栏
aimingoo的专栏
H
Hackread – Cybersecurity News, Data Breaches, AI and More
Microsoft Security Blog
Microsoft Security Blog
博客园 - 叶小钗
腾讯CDC
Engineering at Meta
Engineering at Meta
博客园 - Franky
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
Y
Y Combinator Blog
Recent Announcements
Recent Announcements
Jina AI
Jina AI
A
About on SuperTechFans

Machines – Silicon Republic

Why we have to start thinking ahead to combat rogue AI Opinion: Why dismissing AI doom talk as hype may be the laziest take Microsoft drafts 'humanist AI code of conduct' amid safety debate Quantum funding only a part of the solution, EU experts say OpenAI launches GPT-6 Astra Anthropic launches Claude Fable 5.1 and Mythos 5.1 Household names co-sign open letter on AI cybersecurity threat Bill Gates warns of AI-led societal upheaval if leaders don’t act Revolut to advance fintech architecture with research platform AI bias isn’t just algorithm error, it’s a chain of human decisions Waymo details custom AI silicon for autonomous vehicles OpenAI launches ChatGPT for teens amid calls for stronger child safety measures Alibaba AI models pass Google, Meta with 3bn downloads China’s Z.ai unveils GLM-5.3, claims chart-leading scores Alice & Bob partner on European doctoral network for quantum error correction Quantinuum and Oracle to blend quantum and cloud in new partnership Anthropic pledges to watermark AI content in EU Opinion: AI-native engineering teams are replacing traditional delivery Anthropic confirms plans to build own chips amid global shortage TechIreland launches Galway-bound National AI Challenge 2026 Seven experts join AI Office of Ireland board Thinking Machines Lab unveils new Inkling version at 25pc size After OpenAI incident, Anthropic finds Claude hacked organisations Quantum computing company Zuriq to scale tech amid $25.5m raise Nvidia forms industry alliance post Hugging Face cyber incident Anthropic breaks from AI rivals in open-weight model debate Why is China cracking down on AI-powered companions? Moonshot AI readies to release Kimi K3 for public download J&J table-integrated surgical robot gets FDA approval Samsung’s new robotics division appoints former Boston Dynamics lead
Microsoft releases foundational AI models targeting enter...
Suhasini Srinivasaragavan · 2026-04-03 · via Machines – Silicon Republic

Microsoft wants to offer the ‘most complete AI and app agent factory’.

Microsoft has released three new AI foundational models, created in-house, in a move that places the company in direct competition with enterprise AI rivals, despite its deep ties with OpenAI.

The new foundational models target three of the most commercially viable modalities: transcription, voice and images. The models are already powering Microsoft’s products, including Copilot, Bing and Azure Speech, the company said, and will be available in a preview via the Microsoft Foundry and MAI Playground.

With this, Microsoft is furthering its goals of delivering “the most complete AI and app agent factory”, it said.

‘MAI-Transcribe-1’ is a first-generation speech recognition model expected to deliver “enterprise-grade accuracy” across 25 languages at around 50pc lower GPU costs than its alternatives. The model scores lower than 4pc average ‘word error rate’ on accuracy benchmarks, while GPT-Transcribe is at 4.2pc and Gemini 3.1 Flash is at 4.9pc.

‘MAI-Voice-1’ is a speech generation model that, according to Microsoft, can produce 60 seconds of expressive audio in under one second on a single GPU.

Together, the two models are meant to deliver an audio AI stack capable of assisting in call-centre workflows and other voice-driven services, such as providing live captioning, automatic subtitling and converting interactions into structured data for research.

Microsoft’s second-generation image model, ‘MAI-Image-2’, is expected to offer artists a way to “explore” different visual directions. The model is created in “close collaboration” with artists, the company said, and is meant to help enterprises create branding and communication material.

MAI-Image-2 debuted in third spot on the Arena.ai leaderboard for image model families, and is currently ranked fifth.

Microsoft, valued at $2.7trn, already offers several AI-embedded apps and platform services. Its Copilot Studio lets users build agents, while the Foundry services offer a place to train and scale models.

Meanwhile, a recently announced Copilot integration with Anthropic’s Claude Cowork is meant to target the growing demand for autonomous agents.

Microsoft backed OpenAI in its recent $122bn funding round alongside the likes of Amazon, Nvidia and SoftBank. Late last year, the company announced a $10bn investment plan for a data centre in Portugal. It also announced a $37.5bn quarterly capital expenditure bill at the end of January.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.