Microsoft has unveiled two groundbreaking in-house AI models that signal a dramatic shift in its approach to artificial intelligence infrastructure. The tech giant released MAI-Image-2.5-Pro, its most advanced image generator yet, and MAI-Voice-2-Flash, a speech model designed for high-volume enterprise applications. What makes this launch particularly significant are the production metrics Microsoft shared, demonstrating cost reductions of up to 89% compared to using OpenAI's models across its product suite.
The two models represent opposite ends of the quality-speed-cost spectrum. MAI-Image-2.5-Pro targets premium applications requiring high-fidelity imagery, detailed editing, and precise text rendering within images—a historically challenging task for AI. Already ranking second on Arena's image editing leaderboard, the model is drawing attention from major creative industry players. Meanwhile, MAI-Voice-2-Flash focuses on efficiency, running twice as fast as its predecessor whilst costing 32% less. It's designed for high-volume voice applications like call centres and customer service operations where speed and cost matter more than marginal quality improvements.
Perhaps more telling than the model launches themselves are Microsoft's deployment statistics. Bing Image Creator now runs entirely on MAI-Image-2.5, marking the first time the tool is fully powered by in-house technology. In PowerPoint, the image model reduces GPU costs by 84% compared to OpenAI's GPT-Image-2. OneDrive users are experiencing 26% higher save rates, 25% lower latency, and 2.5 times greater efficiency. The voice model powers Dynamics 365 Contact Center, serving clients like T-Mobile and EasyJet, with reported GPU cost reductions reaching 89%.
Microsoft's healthcare implementation demonstrates the models' real-world impact. Dragon Copilot, used by 170,000 medical providers processing 28 million patient encounters quarterly, now operates on MAI-Transcribe-1.5 across 58 languages. Internal evaluations show a 50% reduction in transcription and language-identification errors—crucial in a field where mistakes can affect patient care. The company's MAI-Code-1-Flash model, deployed in GitHub Copilot, achieves 10% higher code acceptance rates than GPT-5.4 Mini whilst using fewer tokens, with developers showing 6-11% higher retention rates.
CEO Satya Nadella framed the strategy as 'Frontier Diffusion,' arguing that capabilities considered cutting-edge a year ago are now commodities that Microsoft can replicate efficiently for specific, repetitive tasks. He emphasised that whilst frontier models from OpenAI and Anthropic remain part of Microsoft's ecosystem, the company is increasingly routing traffic to its own models when they match or exceed alternatives. This represents a fundamental rethinking of Microsoft's relationship with OpenAI, following reports that their exclusive licensing arrangement has become non-exclusive.
The business logic is compelling: Microsoft claims software now has 'real marginal cost for the first time' due to AI, making optimisation critical when features run on every keystroke across billions of users. An 84% GPU cost reduction isn't merely an improvement—it's the difference between profitability and unsustainable expenses. Microsoft is now packaging its 'hill-climbing' methodology through Azure Foundry, allowing enterprises to train specialised models against proprietary evaluations, transforming an internal cost-cutting exercise into a commercial product.
Fuente Original: https://venturebeat.com/infrastructure/microsoft-launches-new-in-house-ai-models-it-says-cut-costs-up-to-89-versus-openai
Artículos relacionados de LaRebelión:
- IA de OpenAI HACKEA Hugging Face El Fin de la Seguridad Digital
- IA de OpenAI se descontrola y ataca Hugging Face
- AI Models Escaped OpenAIs Security Scare
- IA de OpenAI Hackea Sistemas Autonomamente
- OpenAI Models Escaped Containment and Attacked Hugging Face
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario