Microsoft models in Australia
Alongside its partnership with OpenAI, Microsoft develops its own models. The Phi family consists of small language models released with open weights under the MIT licence, designed to run efficiently on laptops, phones and modest servers. Microsoft AI's MAI models cover reasoning, voice and image generation, and increasingly power Copilot features.
Most Microsoft 365 Copilot users never choose a model directly, because Copilot selects from Microsoft's and partners' models behind the scenes. Developers access Phi and MAI models through Microsoft Foundry, and can download Phi weights to run locally or on their own servers.
Last reviewed October 2026
Strengths
- +Phi models are small enough to run on-device or on modest servers
- +Phi weights are released under the permissive MIT licence
- +Integrated with Microsoft Foundry, Azure security and Microsoft 365
- +Phi models are available in local tools such as Ollama and LM Studio
Limitations
- –Serverless Phi and MAI deployments in Microsoft Foundry are Global Standard, so processing may occur offshore
- –Phi models are less capable than large frontier models on complex reasoning
- –MAI models are closed and available only through Microsoft services
Ways to access these models
Microsoft Foundry
Phi and MAI models as serverless deployments, or Phi on managed compute you control.
Microsoft 365 Copilot and Copilot apps
Microsoft and partner models selected automatically inside Copilot experiences.
Self-hosted Phi
Download MIT-licensed Phi weights and run them locally or on Australian servers.
Keeping data in Australia
Microsoft Foundry's region availability tables list Phi and MAI models in Australia East only as Global Standard serverless deployments, which may process requests in any Azure region. To keep Phi processing onshore, self-host the MIT-licensed weights on your own hardware, on Azure virtual machines in Australia East or Southeast, or on another Australian cloud region. MAI models are closed, so there is no self-hosted alternative. Microsoft 365 Copilot follows Microsoft's own data residency commitments for Microsoft 365 rather than Foundry deployment types; review Microsoft's documentation for Copilot processing in Australia separately. Confirm Foundry options in the region availability tables for Foundry Models sold by Azure, filtering for Australia East and the Standard deployment type.
For workloads where data must stay onshore end to end, see private AI hosted in Australia.
Best for
- →On-device or edge AI where a small model is enough
- →Self-hosted classification, extraction and summarisation on modest hardware
- →Organisations standardised on Microsoft 365 and Azure
Tools built on Microsoft models
General Assistant
Microsoft's AI assistant for work, built into Word, Excel, PowerPoint, Outlook and Teams and available through the Copilot app. It drafts, summarises and analyses using the emails, files, chats and meetings each user can already access, and includes agents such as Researcher and Analyst for deeper tasks. Microsoft now brands the app simply as Microsoft Copilot.
Best for: Organisations already standardised on Microsoft 365 Business or Enterprise plans
Microsoft 365 Copilot review →Automation
Microsoft's low-code platform for building AI agents and automated agent flows. Makers combine generative answers, topics, actions and Power Platform connectors to ground agents in business data, then publish them to Teams, websites, Microsoft 365 Copilot and other channels. Professional developers can extend agents with custom code and APIs.
Best for: Microsoft-centred organisations wanting agents inside Teams and Microsoft 365
Microsoft Copilot Studio review →GitHub Copilot
Free plan + paidCoding Assistant
GitHub's AI coding assistant. It suggests code as you type, answers questions in a chat panel, makes multi-file edits, reviews pull requests, and can take on issues as a coding agent that opens pull requests for review.
Best for: Development teams whose code and workflows already live on GitHub
GitHub Copilot review →Ollama
Free plan + paidDev Tools
An open-source tool for downloading and running open-weight language models on a Mac, Windows or Linux machine, or on your own server. It provides a simple command line, a desktop app and a local API, plus optional paid cloud models for larger workloads.
Best for: Developers prototyping with open models on their own machines
Ollama review →LM Studio
FreeDev Tools
A desktop app for Mac, Windows and Linux that lets you discover, download and chat with open-weight models offline. It includes document chat, a local server with OpenAI-compatible endpoints for developers, and support for MCP tools, all through a graphical interface.
Best for: Non-technical staff who want private AI on their laptop
LM Studio review →
Frequently asked questions
What are Microsoft's Phi models?
Phi is Microsoft's family of small language models, trained to perform well relative to their size. They are released under the MIT licence, so businesses can run them on laptops, phones or servers, and fine-tune them for specific tasks.
What are MAI models?
MAI models are developed by Microsoft AI and cover areas such as reasoning, voice and image generation. They are closed models used in Microsoft products such as Copilot, and some are offered to developers through Microsoft Foundry.
Can Phi models run in Australia?
Yes, if you host them yourself. Phi weights can run on servers in an Australian data centre or Azure virtual machines in Australian regions. Foundry's serverless Phi deployments are Global Standard, which does not keep processing onshore.
Which model does Microsoft 365 Copilot use?
Copilot draws on a mix of models, including OpenAI GPT models, Anthropic Claude models in some features, and Microsoft's own models, and Microsoft changes the mix over time. Check the Microsoft 365 Copilot documentation and your tenant's settings for current model options and data handling.
Other model providers
Building on Microsoft models?
AI Lab Australia designs and builds AI systems on the right model and hosting route for your privacy and quality requirements.