Microsoft Introduces CoreAI and Makes Phi-4 Free for Everyone

So Microsoft is making some big moves with its AI strategy, reshuffling its internal structure and releasing an impressive new open source model. On hugging face, we'll break down the key announcements from Satya Nadella and the team dive into the new core AI platform and Tools division and cover the news about Microsoft's fifor model going fully open source. There's a lot to unpack, so let's get into it. First, let's talk about the reorganization. If you've followed Microsoft for a while, you know they've been pivoting toward AI for the past few years. But in early 2025, Satya Nadella, who's the chairman and CEO of Microsoft, spelled it out bluntly. We're entering the next innings of this AI platform shift and everything's about to move at lightning speed, like 30 years of tech change condensed into just three. So he's basically saying that in 2025, every single layer of the application stack is going to be impacted by AI. It'll be akin to someone inventing a new graphical user interface, plus new Internet servers, plus cloud native databases, all rolled out almost at once. So to keep up with this paradigm shift, Microsoft's forming a brand new engineering group called Core AI Platform and Tools. This group blends together several crucial parts of the company this their developer division, which handles Visual Studio, GitHub and so on. The AI platform team, think Azure, AI Foundry, and certain teams from the Office of the CTO, including folks working on AI supercomputer projects, AI agentic runtimes, and something called Engineering Thrive. Bringing all these teams under one roof means Microsoft can build an end to end Copilot and AI Stack that both Microsoft itself and external customers can use. Essentially, we're talking about the creation of agentic apps, apps with memory entitlements and what they call action space, all powered by big AI models. Jay Parikh, formerly an engineering chief over at Meta, is the new executive vice president in charge of this core AI platform and Tools division. You might recall that Microsoft hired him back in October and and this is his first big splash since joining, he'll have people like Eric Boyd, head of AI Platform, Jason Taylor, Deputy CTO of AI Infrastructure, Julia Lyonson, the leader for Microsoft's developer division, and Tim Bozarth, leading Developer Infrastructure, all reporting to him. This shift really signals that the entire developer division, which used to revolve around traditional coding tools, is now pivoting firmly into AI. In fact, part of their mission is to power GitHub Copilot, which is already a big product that helps developers with AI assisted coding. By merging GitHub Copilot's team with the main AI platform teams, Microsoft wants a tighter feedback loop. Essentially, if something's working in Copilot, they can funnel that insight back into the AI stack for other applications. Satya Nadella also emphasized that Azure is going to become the infrastructure for AI and then everything else like Azure AI foundry, GitHub VS code will be built. On top of that, Microsoft's broader vision is an AI first app stack that changes how every SaaS application is developed. We're seeing new runtimes for agent orchestration, reimagined management layers, and next gen user interfaces that revolve around AI and chat interaction. Developer productivity is going to look totally different, and that's why they've chosen to unify all these divisions under one big AI umbrella. Now shifting gears a bit, Microsoft also made waves by open sourcing a powerful language model called Phi 4. It first appeared last month, but back then it was only available through Microsoft's Azure AI Foundry platform. Now, however, the they've dropped the entire model weights and all onto hugging face under the MIT license. That's a really big deal if you follow open source AI, because most large scale models out there either aren't fully open or come with restrictive licenses. 5 4, on the other hand, is fully out there and it's apparently smashing a bunch of benchmarks, even outperforming some bigger name models in areas like advanced math and coding. So what Sports Special about 5 4? For starters, it's a 14 billion parameter model, much smaller than those massive 70B or 100B giants that was trained on 9.8 trillion tokens of both curated and synthetic data. Microsoft reports top tier results on tests like math, mgsm, human eval and other advanced reasoning challenges. They even say 54 sometimes beats GPT4O, a variant of GPT4O on graduate level STEM tasks. One factor that likely helps is fifor's specialized math training to handle step by step problems more reliably. They also expanded its context window to 16,000 tokens so it can handle longer text inputs like extensive code files or research articles. 54 is a dense decode only transformer, but Microsoft put tons of effort into cleaning and generating its training data. They used synthetic data generation plus curated organic data and finished it off with advanced post training methods, supervised fine tuning, direct preference optimization, and something called pivotal token search. Essentially, they identify the specific tokens that make or break a solution, then teach the model to pick the right ones at those critical moments. Another major concern is hallucination when the model fabricates fact. To tackle this, Microsoft built in a refusal mechanism. If the model isn't sure about an answer, it'll hold back rather than guess. They tested this on a data set called simple qa. Think obscure trivia and the new approach leads to more I'm not sure responses instead of random misinformation. Comparisons to other big name models think Google Meta and thropic are inevitable. Yes, 14B parameters is smaller, but Microsoft says V4 can hold its own or even outperform larger alternatives in certain niches like coding and higher level math. The size difference also means less overhead. Great if you don't have massive GPU farms. And because Fee four is full fully open sourced under mit, you can adapt or fine tune it without license worries. In their technical report, Microsoft notes it excelled on fresh math competitions like the November 2024AMC. To confirm it's not just memorizing old data, they also tested coding tasks such as Human Eval plus dot. If you're a developer, a fast, relatively compact model that still knocks out tough problems is pretty appealing. Still, it's not perfect. 54 can stumble on extremely detailed instructions or less common facts. Microsoft also acknowledges the persistent risks of bias or harmful outputs even after repeated red teaming tests. No generative AI system is foolproof after all. This release ties into Microsoft's broader AI push. Under the new Core AI Platform and Tools group led by Jay Paric, Microsoft is unifying its developer tools and AI infrastructure. They're aiming for an end to end copilot and AI stack, and fifor is part of that strategy. Smaller but mighty models that are fully open. The Satya Nadella wants Microsoft's internal boundaries to vanish in the eyes of users, so everything is integrated under a one Microsoft banner. The idea is to make agentic AI apps as routine as opening Visual Studio once was. So that's the story. A big reorg at Microsoft all about AI plus this major open source release of Fi4, a 14B per megapro model you can now freely explore on hugging face. If you do check it out, just be mindful of its limitations, especially for high stakes use. But hey, that caution applies to any AI tool these days. What do you think will combining dev and AI teams at Microsoft Supercharge innovation is open sourcing top tier models the best path for AI? Let me know your thoughts and if you enjoyed this overview, feel free to share it with anyone who loves AI. Thanks for watching and I'll see you in the next one.