Nextcloud Hub 26 Winter for AI admins: what changed, and an upgrade checklist
· By Bee Flow · 4 min read
Nextcloud 33 promises faster local AI and five to eight times faster transcription. Where those gains come from according to the changelogs, what changed for the Assistant, Context Chat and Talk, and seven checks before you upgrade.

Nextcloud released Hub 26 Winter on 18 February 2026. Under the hood that is Nextcloud server 33, as the maintenance and release schedule shows. The press release leads with ADA, a rewrite of the file access layer, and gives two numbers for AI: "faster responses in local AI apps, saving up to 10 seconds per prompt" and "five to eight times faster local AI transcription".
If you run the Assistant on your own hardware, those two lines are the reason to read on. Below is what the release post, the app changelogs and the admin manual say changed for AI, followed by the checks we would make before upgrading an instance that has AI apps on it.
Where the speed comes from
The launch blog lists the AI figures under "Faster apps", next to Photos and Talk. It does not say which change produces them, and the ADA engine post of 4 February is about file access in general. So we went looking in the changelogs.
Transcription is the easy one. The stt_whisper2 changelog for version 2.4.0, dated 3 December 2025, says the app now "ships with Whisper large-v3-turbo model (similar quality as large-v3, but 5-8 times faster)". That is the same range as the press release. The gain arrives with the speech-to-text ExApp, so you get it when stt_whisper2 is updated.
For prompt latency, one server change points the same way. Nextcloud 33 adds triggerable task processing providers: when a new task arrives and nothing else is running, the server nudges the ExApp to fetch it, instead of the task waiting for the app's next poll. The pull request calls this part of a project "to reduce the polling pressure on the server from ex apps", and the stt_whisper2 2.3.0 entry describes its support for the trigger event as a way "to speed up task pickup". The llm2 changelog mentions polling performance improvements in 2.5.0. Nextcloud has not tied the 10-second figure to any one change, so read this as our interpretation.
Assistant, Context Chat and Talk
The Assistant gets memories. In the launch blog's words, "it can now remember what you talked about across all your chats", and users can remove what it keeps. The Assistant changelog shows memories landing in 3.1.0 on 15 January. Users can also label AI-generated images, documents and audio, which Nextcloud says "aligns with the EU AI Act", and the feature list names watermarking for AI-generated content.
On models: llm2 now ships Olmo 3, the IBM watsonx integration supports self-hosted watsonx, and the OpenAI and LocalAI integration can connect to more than one service, with improvements to quotas and reporting. The mobile apps get a chat with the Assistant.
Context Chat gets a single line in the feature list: "added possibility to scale the embedding process". The Context Chat page in the admin manual shows how. You point the backend at an OpenAI-compatible embedding service with CC_EM_BASE_URL, plus variables for model name and credentials, in the deploy options.
In Talk, live transcription now comes with translation. The live_transcription page lists Nextcloud 33 as the minimum version. It needs the High-Performance Backend (released after September 2025), a deploy daemon in AppAPI, and a translation provider such as translate2 for the translation part. The two connection variables, LT_HPB_URL and LT_INTERNAL_SECRET, have to be set at install time.
Flow, the Windmill ExApp, is not part of the release post. Its changelog added Nextcloud 33 support in 1.2.2 in December and moved to a much newer Windmill in 1.3.0 in January.
Before you upgrade
This is the list we would work through on an instance with AI apps.
- Check PHP and the database. The upgrade notes for 33 drop PHP 8.1, deprecate 8.2 and add 8.5. PostgreSQL 13 and Oracle 11g are no longer supported for Nextcloud's own database.
- If you restrict outbound traffic, allow
connectivity.nextcloud.com, which now replaces www.nextcloud.com for the connectivity check. Outgoing requests also carry a new user agent,Nextcloud-Server-Crawler/X.Y.Z, which matters if a proxy rule matches the old one. - Set the server ID for the new Snowflake IDs in
config.phpor through environment variables, as the same notes ask. - Update the AI ExApps together with the server, and check that each one you run lists support for 33. For Context Chat, the manual says the major and minor versions of
context_chatandcontext_chat_backendmust match. Update both. - Restart any dedicated AI workers. If you run task processing workers as described under "Improve AI task pickup speed" in the AI overview, the manual notes that a running worker does not see changed PHP code until it restarts.
- After the upgrade, read the admin overview. Nextcloud 33 adds a setup check that tells you when AI tasks fail, so a provider that needs attention shows up there. On HaRP setups, AppAPI for 33 registers HaRP automatically once it is installed.
- Optional: the new
/metricsendpoint speaks OpenMetrics and answers only on localhost by default. It is a cheap way to watch the instance while the workers are busy.
Our own advice is to upgrade a test copy first and time a few prompts and one transcription before and after. Nextcloud gives the figures as "up to", and the result depends on your hardware and on which apps you run.
Where Bee Flow fits
Some teams want more next to the Assistant: chats with a choice of models, cloud or EU-hosted, knowledge bases whose answers cite the document they came from, and a check that stops personal data before a message leaves. Bee Flow is a self-hosted AI workspace that does these things on a server of your own, beside your Nextcloud. It adds to the Assistant, and the two can run side by side. The Nextcloud page explains how they fit together. If you want to talk through your own setup, get in touch.