AI Issue: Microsoft's Own AI Model Preview
Microsoft AI has unveiled MAI-Voice-1, a system that expresses emotions through voice, and MAI-1-preview, a massive model trained on 15,000 GPUs.
MAI-Voice-1 for emotionally expressive voice and MAI-1-preview for intelligent text processing are being showcased for the first time in Copilot.
Microsoft AI's New In-House Models
- Microsoft AI (MAI) has released two in-house models.
- MAI-Voice-1 for natural voice generation and MAI-1-preview, the first self-developed foundation model.
- MAI-Voice-1 delivers high-quality, emotionally expressive audio at an impressive speed, generating 1 minute of audio in under 1 second on a single GPU.
- The voice model is already being used in Copilot Daily and Podcasts, and can be tested in Copilot Labs.
- MAI-1-preview, a mixture-of-experts foundation model, was trained on approximately 15,000 NVIDIA H100 GPUs and is currently undergoing public testing on LMArena.
- The foundation model will be gradually rolled out to Copilot text features, with API access for trusted testers.
Microsoft AI's Goal: AI for Everyone
Microsoft AI (MAI) aims to create AI that helps every individual and organization reach their full potential.
Microsoft envisions AI as a helpful and trustworthy companion -- a gateway to knowledge and a range of capabilities tailored to people's specific needs.
To realize this vision, MAI has been building purpose-built models through world-class teams and facilities.
This week marks the first preview of two in-house systems built to advance that goal.
MAI-Voice-1: High-Speed Emotionally Expressive Voice Generation
The first release is MAI-Voice-1, a voice generation model designed to create natural, emotionally expressive, high-quality audio for single or multi-speaker scenarios.
Performance: MAI-Voice-1 can generate 1 minute of audio in under 1 second on a single GPU, making it one of the most efficient voice systems available today.
Usage: It's already integrated into Copilot Daily and Podcasts, providing more natural audio for these features.
Experience: The model is also available in Copilot Labs, where users can test demos like storytelling and guided meditation created with simple inputs.
MAI-Voice-1 makes voice faster and more emotionally expressive, positioning voice as a primary interface for future AI companions.
MAI-1-Preview: A Foundation Model Trained on 15,000 GPUs
The second major achievement is MAI-1-preview, the first foundation model trained entirely in-house from start to finish.
This model follows a mixture-of-experts architecture and underwent both pre-training and post-training on approximately 15,000 NVIDIA H100 GPUs.
Evaluation: The model is undergoing public testing on LMArena, a community platform for model evaluation.
Use Cases: Built to follow instructions and handle useful everyday responses, it will be applied to select Copilot text features over the coming weeks.
Access: Beyond LMArena, trusted testers can access the model through API testing, enabling Microsoft AI to gather targeted feedback.
Access can be requested here.
This marks the beginning of MAI's strategy to deliver improved in-house foundation models while leveraging new approaches from partner models and open source to ensure the best results across products.
Future Plans: Specialized Models for Various Use Cases
Microsoft AI emphasizes that these two models are just the first step of a larger strategy.
Beyond foundation systems, Microsoft plans to create a variety of specialized models tailored to specific user intents and contexts.
This approach is designed to deliver greater value to customers and ensure that Copilot and other Microsoft products can adapt to the millions of diverse interactions they help with every day.
Q&A
Q: What models has Microsoft AI released?
A: Microsoft AI (MAI) has announced MAI-Voice-1, an emotionally expressive voice generation model, and MAI-1-preview, its first foundation model.
Q: What makes MAI-Voice-1 special?
A: MAI-Voice-1 generates natural, high-quality audio at impressive speed, producing 1 minute of voice in under 1 second on a single GPU.
Q: Where can MAI-Voice-1 be used?
A: It's already integrated into Copilot Daily and Podcasts, and can be tested through Copilot Labs which showcases storytelling and meditation guidance demos.
Q: How was MAI-1-preview trained?
A: MAI-1-preview is a mixture-of-experts foundation model trained on approximately 15,000 NVIDIA H100 GPUs, designed to follow instructions and provide useful responses.
Q: How can developers test MAI-1-preview?
A: It's publicly available on LMArena for open evaluation, with additional access available through API testing for trusted users.
Implications
The release of MAI-Voice-1 and MAI-1-preview demonstrates Microsoft AI's effort to build its own core models while leveraging new approaches from partnerships and open source.
Many observers see this as Microsoft taking a step toward greater independence from OpenAI products while investing in developing its own large-scale systems for the future.
For users, this means more emotionally expressive, human-like interactions through voice, and access to more capable, responsive models for text-based use in Copilot.
For the industry, it showcases Microsoft's strategy of combining general-purpose foundation models with specialized systems to meet diverse user needs.
Above all, Microsoft AI is advancing toward a long-term vision where voice becomes a core interface for AI companions and foundation models become the backbone of reliable, practical AI.
These early in-house releases mark the beginning of a larger portfolio designed to deliver stable, independent AI for everyday life.
IMPAKERS Blog | Apple Gemini Siri Integration Issue Read More
Source: Alicia Shapiro, AiNews, "Microsoft AI Introduces MAI-Voice-1 and MAI-1-Preview Foundation Model", https://www.ainews.com/p/microsoft-ai-introduces-mai-voice-1-and-mai-1-preview-foundation-model, (2025-08-29)