For enterprises, solutions like RHEL AI<\/a> and SUSE Linux Enterprise Server<\/a> have been catering to organizations that want AI woven into their infrastructure.<\/p>\n Now, it looks like Canonical<\/a> is jumping onto the bandwagon, as Ubuntu moves towards AI<\/a>, and before you start calling it Ubuslop<\/em> or something along those lines, understand how they are going about this.<\/p>\n Jon Seager<\/a>, VP of Engineering at Canonical, has published a post on Ubuntu Discourse laying out the company’s AI roadmap. The short version is that AI is coming to Ubuntu<\/strong>; it will be local-first<\/strong>, and it will be built around open-weight models<\/strong> and open source tooling<\/strong>.<\/p>\n He laid out a framework distinguishing between two kinds of AI features, implicit<\/em> and explicit<\/em>. Implicit AI is about making existing OS features smarter in the background without requiring users to learn anything new or interact with anything that looks like AI.<\/p>\n He gave the examples of speech-to-text and text-to-speech, both of which can be improved using local inference<\/a> with open source inference tools and open weight models, running entirely on-device.<\/p>\n Explicit AI features are a different story. These are the more obviously AI-centric, agentic workflows that could automate troubleshooting, create documents or applications, and run scheduled maintenance on a fleet of machines.<\/p>\n Jon also gave everyone a look at what it could look like in practice:<\/p>\n Imagine being able to ask your Linux machine to troubleshoot a Wi-Fi connection issue, or to stand up an open source software forge that\u2019s pre-configured, secured, and reachable over TLS.<\/p><\/blockquote>\n One could easily imagine using such a capability as a gateway for controlling your Linux machine from other devices through a variety of mediums – be that a mobile app, text messaging, voice commands or otherwise.<\/p><\/blockquote>\n The delivery mechanism for all of this is inference snaps<\/a>. Rather than asking users to wrangle separate tools, sift through Hugging Face<\/a>, and figure out which model format works on their hardware, Canonical wants a simple And since snaps<\/a> carry the same confinement<\/a> rules as everything else in the ecosystem, the models are sandboxed and cannot freely reach into your files or data.<\/p>\n The local inference approach is what makes this worth paying attention to. The default is not a cloud call to some API that logs your prompts and charges you per token. It runs on your hardware, stays there, and does not require signing up for anything<\/strong>.<\/p>\n Of course, cloud and external services are still an option<\/strong>, but only as a fallback for people who specifically need them, not the assumed path. That is a bigger deal than it sounds, btw.<\/p>\n Most AI integration announcements from Big Tech<\/a> players start from the opposite assumption\u2014cloud first, local maybe someday<\/em>.<\/p>\nWhat’s happening?<\/h2>\n
![\"cropped](\"https:\/\/itsfoss.com\/content\/images\/2026\/04\/jon-seager-ubuntu-ai-roadmap.png\")
<\/figure>\nsnap install<\/code> to handle everything, with hardware-optimized builds served based on your silicon.<\/p>\nMakes sense, I guess?<\/h2>\n
Should you be worried?<\/h2>\n
![](\"https:\/\/feed.itsfoss.com\/link\/24361\/17326955.gif\")
<\/p>","protected":false},"excerpt":{"rendered":"