Canonical plans to introduce AI-backed features as a strictly opt-in preview in Ubuntu 26.10, using local models by default, said Jon Seager, the company’s vice president of engineering. The proposed additions include accessibility improvements and agent-based workflows for tasks such as troubleshooting and creating documents.
Seager said Ubuntu 26.04 LTS does not ship these planned integrations, which remain to be developed as Canonical works toward 26.10. He said AI features would arrive as they reached sufficient maturity and quality.
User choice and local processing
Seager said the tools would default to running inference locally on the user’s machine. Using cloud-based inference would require users to configure it explicitly and provide an API token or another credential. He also said sending logs to the cloud was not part of the plans.
For subsequent releases, Seager proposed a step in the initial setup wizard allowing users to choose whether to enable AI-native features. He said the size of large language models would prevent their inclusion in the installer, so users declining those features would not have the models installed at first run.
Canonical does not plan to add a single global AI switch, Seager said. Instead, the capabilities would be delivered as removable snap packages layered on top of the existing Ubuntu system. Users would be able to remove the relevant packages to remove the planned features.
Seager said inference snaps are subject to the same confinement rules as other snaps, presenting this as a way to restrict access to machines and data. The degree of isolation depends on a package’s confinement level and permissions: strictly confined snaps use interfaces to access additional resources, while classic snaps have access comparable to traditional software packages.
Accessibility and task assistance
Seager divided the roadmap into two categories. The first would use models to improve existing operating-system functions without requiring users to adopt new workflows. He identified speech-to-text and text-to-speech as accessibility functions that could benefit from local inference.
The second category would introduce more explicit AI capabilities, including agents that could help author documents or applications, automate troubleshooting and perform personal automation tasks. These are proposed uses rather than a confirmed feature list for Ubuntu 26.10.
Examples Seager described included diagnosing a Wi-Fi connection problem and setting up an open-source software forge secured and reachable over TLS. He also discussed potential server-administration uses, such as interpreting incident logs and performing scheduled maintenance.
For those workflows, Seager said his aim was to support existing operational boundaries through read-only analysis, narrowly scoped permissions for actions and auditability of decisions and outcomes. He identified security and confinement controls as necessary to prevent unwanted side effects.
Hardware and model limits
The plans build on Canonical’s inference snaps. The company says these packages simplify local model deployment by automatically selecting inference engines and model variants suited to a device’s hardware, where the relevant optimizations are available from silicon partners.
Seager acknowledged that local inference depends on capable hardware and that smaller models cannot yet match larger models for many tasks. He said Canonical planned to expand the range of hardware-optimized variants available through inference snaps.
Canonical’s stated preference is for open-weight models paired with open-source software for running them. Seager distinguished access to model weights from the transparency traditionally associated with open-source software, saying model selection would also consider licensing terms.
Within Canonical, Seager said engineers were being encouraged to experiment with AI tools without targets for token consumption or the proportion of code written with AI. He said engineers would continue to be assessed on their delivery rather than how much they used AI, and that AI-assisted work should be controlled and reviewed. He also rejected careless automated contributions to open-source projects.
