Under the hood of Microsoft AI models | DEM323

Name: Under the hood of Microsoft AI models | DEM323
Uploaded: 2026-06-03T14:08:28+00:00
Description: Dave Citron (CVP, Microsoft AI) walks through what goes into training Microsoft’s latest MAI model family—covering new thinking, coding, voice,...

Today by Dave Citron

Dave Citron (CVP, Microsoft AI) walks through what goes into training Microsoft’s latest MAI model family—covering new thinking, coding, voice, transcription, and image models, plus the architectural and evaluation choices behind their capabilities and performance.

Overview

Microsoft AI (MAI) presents a technical, “under the hood” look at a newly announced family of Microsoft models, including:

Thinking models
Coding models
Voice models
Transcription models
Image models

The session focuses on what it takes to train these models, what the team learns during training and evaluation, and how those learnings are reflected in model architectures, features, and capabilities.

What’s covered (from the session description and chapter list)

Model family announcements and highlights

Introduction of Transcribe 1.5, Voice 2, and Code 1 Flash performance highlights.

Voice 2: speech generation focus

Launch of Voice 2 with emphasis on natural prosody.
Showcase of fine-grained emotional control and multilingual availability.
Demonstration of an emotional tone example (“joy”).

Evaluation and reporting

Reference to a 100-page technical report describing model development.
Discussion of benchmark performance and real-world code evaluation.

Tuning approach

Introduction to Microsoft Frontier Tuning.

Real-world example

A real-world example is referenced: Land O’Lakes quality report generation.