{"slug": "exploring-microsofts-ai-foundation-models", "title": "Exploring Microsoft’s AI foundation models", "summary": "Microsoft has begun rolling out its own foundation models under its Microsoft AI (MAI) program, targeting lower token cost than competitors, with most MAI models available through Azure's Foundry platform, which offers access to more than 11,000 models. Two MAI models are reserved for Microsoft applications: MAI-Cyber-1.1-Flash runs in Microsoft's MDASH security system and MAI-Code-1.1-Flash runs in GitHub Copilot, while MAI-Thinking-1 serves the most complex applications and Microsoft says it runs with a much smaller footprint than competing models. Microsoft also offers a tuning service for building custom versions of the MAI models with customer data, and MAI describes its approach as \"humanist,\" focused on safety and environmental impact.", "body_md": "When Microsoft started to deploy AI models in what is now its Foundry platform, it was building on its partnership with OpenAI. OpenAI was able to use Azure infrastructure to train its models, and in return, Microsoft became the first commercial partner to use them.\n\nAzure OpenAI has been a successful product for both businesses. But as more organizations got access to OpenAI models, Microsoft’s stable of models grew to include not only OpenAI’s, but also models from Anthropic and Hugging Face, as well as models from Meta, Mistral, and DeepSeek. At the same time, Microsoft has continued to offer its own Cognitive Services models for computer vision and text processing, as well as the Phi series of efficient lightweight models for use on the edge.\n\nWith access to [more than 11,000 different models](https://azure.microsoft.com/en-us/products/ai-foundry/models/?msockid=1496141decc46635160801e7ed246782#tabs-pill-bar-oc92d8_tab1), Foundry has proven to be a useful way to build modern AI-based applications, offering many agent-focused features, including tools to ground models using Microsoft’s graph-based IQ data sources. But one key element has been missing — Microsoft’s own large-scale frontier models.\n\nIt’s not that Microsoft is new to developing its own AI models. Microsoft Research has been a hub for [deep learning](https://www.infoworld.com/article/2260824/what-is-deep-learning-algorithms-that-mimic-the-human-brain.html) for much of the last decade, with models that regularly top benchmark tables. But it hasn’t taken the same general-purpose approach as Anthropic, OpenAI, or even Google, instead producing more focused models.\n\nNow Microsoft has begun to roll out its own foundation models as part of its Microsoft AI (MAI) program. With a focus on offering lower token cost than competitors, MAI has spotted a gap in the market. It’s a sensible move to make, too. The days of free and low-cost tokens have gone, never to return, as hyperscale vendors look to recoup the massive multi-billion-dollar investments they’ve made in AI data centers.\n\nSo-called “[tokenmaxxing](https://www.infoworld.com/article/4208123/tokenmaxxing-the-strangest-developer-productivity-metric-of-all-time.html)” is no longer economically viable, dumping massive data sets into LLMs with the hope of getting results. Instead, [inferenc](https://www.infoworld.com/article/4210689/ai-inference-5-best-practices.html)e needs to be done at a lower cost, using models to help choose models and dedicated hardware that can process tokens for fewer watts and less heat. With that philosophy at the heart of its design, [MAI describes its approach as “humanist”](https://microsoft.ai/), with a focus on safety and environmental impact.\n\nMAI models may not have the reach of the more popular general models from Anthropic or OpenAI, but the different models provide support for common scenarios, from cybersecurity, to coding, to transcription. As part of its MAI offering, Microsoft provides a tuning service that allows you to use your own data to build custom versions of the models.\n\nWhile most of [the MAI models](https://microsoft.ai/models/) are available through Azure’s Foundry service, two are reserved for other applications. Currently [MAI-Cyber-1.1-Flash](https://microsoft.ai/models/mai-cyber-1-1-flash/) is used as part of Microsoft’s MDASH security system, and [MAI-Code-1.1-Flash](https://microsoft.ai/models/mai-code-1-flash/) is used as part of the GitHub Copilot service.\n\nThe most complex AI applications can use [MAI’s Thinking model, MAI-Thinking-1](https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-mai-thinking?tabs=python). Microsoft notes that, like other MAI models, MAI-Thinking-1 runs with a much smaller footprint than competing models but still delivers the results users are expecting. Certainly, published benchmarks show good results — and Microsoft isn’t afraid of publishing numbers that favor some of the bigger models from Anthropic and OpenAI.\n\nThe key for Microsoft is to show that it gets satisfactory results at a lower cost. The aim is not to exhaust token budgets, but to gain market share by showing that you can get results that are nearly as good, for a much lower cost. There’s little downside to switching to using the new MAI models, as they can be called using the existing Foundry Azure OpenAI APIs. If you’ve got existing code, start by deploying MAI-Thinking-1 to your Foundry instance, then change your API call to the model’s endpoint URL and start sending messages to the deployed model. You can set a limit on the number of tokens used to complete a request, helping manage your application budget.\n\nMAI-Thinking-1 can use tools via function calls, so you can build it into your own agents. More complex operations can take advantage of its ability to save encrypted reasoning chains from request to request, allowing you to quickly move to the next response in a chain, refining data to produce new prompts as your agent runs.\n\nLike other large models, MAI-Thinking-1 can stream chunks of data back to your code instead of delivering a complete chat object. This can help with user experience, letting users see progress rather than being left waiting for a response with no cues. Of course, if you’re using MAI-Thinking-1 in a long-running workflow, there’s no need to stream data, as you’re passing responses through a chain of requests, possibly using [Microsoft’s new decision-scoring model](https://commandline.microsoft.com/microsoft-decision-1-model-foundry/) as part of an agent router to determine which steps are chosen and which models each step uses.\n\nSome of the more interesting work MAI has been doing recently is starting to ship as new “Flash” models, which work locally or in the cloud. This approach can lead to significant savings, while offering similar fidelity to cloud-only models. At the same time, using local inference on NVIDIA RTX-class hardware uses significantly less power than data center hardware, lowering the environmental impact. Even though Flash models are smaller, they’re described as foundational models.\n\nFlash models [like MAI-Code-1.1-Flash](https://microsoft.ai/models/mai-code-1-flash/) aren’t yet small enough for the relatively low-cost Arm-based hardware that Microsoft has branded as Copilot+ PCs, as they need more than 120GB of RAM to run. However, new hardware like the NVIDIA RTX Spark-powered Surface Ultra, when configured with at least 128GB of RAM, should be more than capable. You will then need to consider the amount of agent usage, to determine whether an expensive device like Surface Ultra makes local operation more economical than in cloud servers.\n\nMAI-Cyber-1.1-Flash has a quite different role, providing offload capabilities to the agents that power the MDASH cybersecurity platform. Here the Flash model is used to manage 90% of the tasks that used to be delivered by much larger and more expensive models. This makes the service more economical to run, without affecting performance. When an agent powered by the Flash model can’t handle an operation, it passes it on to another, more powerful model. The MAI-Cyber-1.1-Flash release also reduces operational costs over the 1.0 version, as it’s 18% cheaper to run.\n\nMicrosoft’s investment in AI is more than the models. The [latest generation of its Azure Maia inference accelerators](https://azure.microsoft.com/en-us/blog/deep-dive-into-azure-maia-300-architecture/) is optimized for Microsoft’s own models as well as for OpenAI. This approach should allow Microsoft to reduce end-user token costs, as the underlying dataflow architecture allows inference software to place data where it is most efficiently accessed by the accelerator.\n\nThe Maia architecture also allows Microsoft to use non-uniform memory access to manage memory usage. It can do this because Maia is no longer simply an accelerator design, but an entire platform from rack to server to networking. This means the software can be designed to take advantage of this new infrastructure element, using what Microsoft calls “[Software Defined Locally Accessed Dataflow Architecture](https://techcommunity.microsoft.com/blog/azureinfrastructureblog/maia-200-software-defined-dataflow-and-all-ethernet-networking-for-efficient-inf/4548198)” (SDLA) and a chiplet-based silicon design that separates memory and I/O functionality.\n\nPrevious versions of Maia were fast, but the Maia 300 release boosts compute operations by 36% with 2.5x improvements in high bandwidth memory performance. This translates into roughly a 2x overall performance improvement per watt. Jobs don’t need to be handled by one accelerator; 96 accelerators are available in each rack. Plus, the Maia architecture supports multiple jobs per rack, sharded across multiple racks. In short, Maia can deliver multi-tenant inference without blocking operations, scaling up and out as necessary.\n\nMicrosoft’s investment in its own foundation models seems to be paying off, buoyed by a focus on the economics of running and using LLMs. Keeping inference costs and power requirements down is a logical path to take — and one that should continue to pay dividends over the long term. You don’t need to deliver the absolute best model; you only need one that’s good enough and cheap enough to use in most applications. That’s what Microsoft is delivering with Maia, and it’s an approach that should appeal to companies looking at growing AI bills.", "url": "https://wpnews.pro/news/exploring-microsofts-ai-foundation-models", "canonical_source": "https://www.infoworld.com/article/4245368/exploring-microsoft-ai-foundation-models.html", "published_at": "2026-11-12 09:00:00+00:00", "updated_at": "2026-11-12 09:23:03.380322+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-infrastructure", "ai-tools"], "entities": ["Microsoft", "Microsoft AI", "Azure", "Foundry", "OpenAI", "MAI-Thinking-1", "MAI-Cyber-1.1-Flash", "GitHub Copilot"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/exploring-microsofts-ai-foundation-models", "markdown": "https://wpnews.pro/news/exploring-microsofts-ai-foundation-models.md", "text": "https://wpnews.pro/news/exploring-microsofts-ai-foundation-models.txt", "jsonld": "https://wpnews.pro/news/exploring-microsofts-ai-foundation-models.jsonld"}}