Microsoft AI has launched seven new artificial intelligence models, expanding its in-house MAI family across image, voice, transcription, coding and reasoning.
The line-up includes MAI-Thinking-1, a reasoning model; MAI-Code-1-Flash, a coding model linked to GitHub Copilot and Visual Studio Code; MAI-Image-2.5 and its Flash version for image generation and editing; MAI Transcribe-1.5 for speech-to-text; and MAI-Voice-2 for speech generation, alongside a lower-cost Voice-2-Flash.
Microsoft AI said the models were developed internally and trained from scratch rather than distilled from third-party systems. It said the family shares a common data, infrastructure and evaluation framework, with products designed to work together across business tasks.
The launch also outlined a broader strategy for adapting models to customer workflows through what Microsoft calls Frontier Tuning. The approach centres on reinforcement learning in real-world environments, allowing organisations to train models on the steps and decisions that define their work processes.
Companies would use their own operational data in private training environments. Microsoft AI said customers would be able to tune model weights themselves, reflecting a push to give developers and businesses more control over how systems are adapted for specific tasks.
Microsoft argued that custom tuning can improve both efficiency and output. It said a tuned MAI model for Excel matched "GPT 5.4" while being up to 10 times more efficient, and that in testing for an unnamed organisation with strict internal standards, an MAI model achieved the highest win rate at about one-tenth the cost of rival models.
Healthcare push
Alongside the launch, Microsoft said it is working with Mayo Clinic on a healthcare AI model based on the clinic's clinical expertise, de-identified clinical data and longitudinal insights.
The model is intended first for use within Mayo Clinic before becoming more widely available through Azure Foundry. Microsoft said Mayo Clinic would own the model, presenting that structure as part of a focus on patient trust, safety and control over clinical data.
The project highlights Microsoft's interest in sectors where data sensitivity and domain knowledge are central to adoption. Healthcare has been one of the most closely watched areas for AI deployment because of the demands around accuracy, oversight and stewardship of patient information.
In-house emphasis
Microsoft AI also described the launch as part of a broader effort to build long-term self-sufficiency in frontier model development. It said it had built every part of the stack itself, from architecture and training pipelines to post-training systems, and that it co-designs models with Maia 200 silicon.
According to Microsoft, those efforts have already delivered a 1.4-times efficiency improvement. It framed that engineering approach as part of what it called a "hill-climbing machine" designed to improve models continuously through more compute, better data and tighter evaluation.
The announcement also underscored how leading AI groups are trying to differentiate not only on model size or benchmark scores, but also on ownership, deployment control and cost. For large customers, especially in regulated industries, those factors are increasingly as important as raw model performance.
Microsoft said the compute used to train frontier models has increased by a factor of one trillion and predicted another thousand-fold increase over the next three years. That expectation underpins its push to expand internal model development at a time of intense competition among major technology groups and specialist AI labs.
It presented the MAI family as a multimodal system that can support real-world workflows across software engineering, speech, images and reasoning. Microsoft said MAI-Thinking-1 matches leading models on key software engineering benchmarks, MAI-Image-2.5 surpasses the Arena score of Nano Banana Pro, MAI Transcribe-1.5 supports domain-specific terminology in 43 languages, and MAI-Voice-2 covers 15 languages.
Microsoft also said the models will be distributed through Foundry and used across its own products, including developer tools. The strategy ties research more closely to the company's software estate, with coding and productivity applications likely to serve as early proving grounds for the new systems.
The group also set out an ambition it called "Humanist Superintelligence", describing advanced AI systems as tools that should remain under human oversight and subordinate to human goals. "People - you - must always remain in control," the MAI team said.