On a new chat, picking Temporary from the / menu did nothing, so the next message was saved and showed up in the chat list. Picking it now turns the chat temporary, the same as the Temporary Chat button in the header.
Fixes#34607
Opening a long PDF in the preview drew every page at full resolution before showing anything and kept them all in memory, so a 1500-page manual took over half a minute, and one report had a 357-page scanned file push Firefox past 20 GB. The preview now lays out all pages right away, draws only the pages near the screen and releases pages once they scroll far away. Jumping to a cited page, zooming and scrolling work as before, and comment highlights and the zoom level apply to far-off pages as they are drawn. The page thumbnails also release each page after drawing it; without that, the scanned file below still passes 9 GB while its thumbnails load. The browser's find (Ctrl+F) now only matches text on pages near the screen, and selecting text across into far-off pages no longer works.
| Large PDF in the file preview | before | after |
| --- | --- | --- |
| 1500-page text PDF, opened at page 1200 | 35.7 s, 7.5 GB | 1.3 s, 1.0 GB |
| 400-page scanned PDF (39 MB), opened at page 300 | nothing shown, over 9.8 GB after 9.4 s (stopped) | 1.4 s, 1.1 GB, 1.5 GB peak scrolling the whole file |
Fixes#29576
Admin settings pages give no hint of what is about to be saved, so an accidental edit goes through unnoticed, for example a number field that changes when you scroll over it. Next to Save, each admin settings page now shows how many settings differ from what was last loaded or saved. Clicking that count or Save lists every changed setting with its old and new value and a Save button to confirm, and fields the page keeps hidden, like API keys and passwords, stay hidden there too. With nothing changed, Save works as before. Connections and Integrations keep their plain Save with no review, controls that already save on their own keep doing so, such as the image and video connection lists or switching the text-to-speech engine (which saves the whole Audio page), and Events, External Knowledge, Models and Evaluations are unchanged. Save now shows the not-allowed cursor while it is disabled during a save, as it already did on Images and Videos.
Discussion: https://github.com/open-webui/open-webui/discussions/34532
* fix: rename Capabilities to Features in the model editor
The model editor section called Capabilities mostly holds switches for which features users see on a model, such as image generation, web search and the code interpreter, so the name led admins to expect it to describe what the model itself can do. The section is now called Features, and its tooltip no longer says the connected model and provider must support every switch, since each switch's own description says what it needs.
* Leave the English value of the new tooltip empty like its neighbours
* fix: a prompt from an embed goes to the wrong chat or to two chats at once
With a note open and its chat beside it, a prompt sent by an embed in that chat was picked up by that chat and by the main chat behind the note tab, so a tool embed's "ask the chat" button opened two confirmation dialogs. In a side chat the same button asked the prompt in the main chat. Each embed now belongs to the chat that shows it, and only that chat takes its prompt; a prompt from anywhere else goes to the page's main chat as before.
Fixes#34152
* fix: route a prompt from any embedded frame to the chat that shows it
The chat a prompt belongs to is now read from where the sending frame sits on the page, for every frame and not only tool embeds, so artifacts and inline HTML in a chat tab or beside a note also reach the right chat. A prompt that comes from no frame goes to the main chat.
* Name the message's source chat apart from the chat object
Exporting a knowledge base packed and compressed all of its files while holding up the worker process serving it, so every other request on that worker (chats, page loads, health checks) waited until the archive was done. With several exports running at once, other users saw pauses of more than a second. The download still waits for the archive, but the server now keeps answering everyone else while it is packed. In our test the worst pause dropped from about 1.3 seconds to about 0.4 seconds, and the downloaded file is unchanged.
| 200 files of ~60 KB, 10 exports at once, worst of 2 runs | before | after |
| --- | --- | --- |
| slowest health check while exporting | 1328 ms | 403 ms |
Fixes#34599
Large profile pictures of users, models and channel webhooks (animated ones, or ones set through the API) were sent out in thousands of small pieces, each handled separately, and the knowledge base export sent its zip file the same way. A 2 MB picture took around 200 to 400 ms to load, 20 loads of it at once, as on a page full of avatars, took 9 to 15 seconds, and an 8 MB export spent close to a second just sending the file. Both are now sent as one complete response with the same content type and headers, plus a size header in place of chunked sending.
| Picture | One load before | One load after | 20 at once before | 20 at once after |
| --- | --- | --- | --- | --- |
| Model, 2 MB | 207 ms | 15 ms | 9.3 s | 0.48 s |
| User, 2 MB | 379 ms | 15 ms | 15.0 s | 0.34 s |
| Model, 5 MB | 743 ms | 39 ms | 33.6 s | 0.63 s |
| User, 5 MB | 627 ms | 25 ms | 30.7 s | 0.54 s |
Fixes#34597
* fix: uploads stall behind a knowledge base sync waiting on a slow extraction engine
Every uploaded file started processing at once, and each file waiting on the extraction engine used up one of the server's limited worker slots, the same slots that save new uploads to storage. With a slow extraction engine a large knowledge base sync used up all of them, so every new upload, chat attachments included, had to wait for an extraction to finish before it was even saved. Files uploaded into a knowledge base now process at most half as many at a time as there are slots, and the rest wait their turn without using one, so other users' knowledge base uploads may wait too during a large sync. Chat attachments and all other uploads skip this line and process right away, as before. Measured with an extraction engine limited to one connection and 50 files synced into a knowledge base, uploads stayed under 0.1 s instead of taking up to 4 s, a chat attachment sent meanwhile was ready in 20 s instead of 42 s, and the sync took the same time.
Related to #34369
* Size the knowledge upload limit inline and type it
* Ignore an invalid thread pool size like the server does
New **pt-BR** translations for items introduced in the latest releases, plus a small consistency pass across existing strings. Placeholders preserved. No logic changes.
Connecting an MCP server over OAuth 2.1 failed with "Algorithm 'HS256' requires 'oct' key" when the server asked for the openid scope and signed the ID token it returned with the client secret, as Readwise does. Open WebUI tried to check that ID token against the server's published keys, which cannot work for this valid signing method, and threw away the access token it had just received. MCP connections only use the access token, so the sign-in no longer checks the ID token and the connection goes through. SSO login is unchanged.
Fixes#34504
With Redis, every large model profile image makes each load of the model list slower, and Redis holds a full copy of every image, even though the list never shows them (the interface fetches each picture separately). Animated or imported profile images are kept at full size and can be several MB each. Open WebUI now leaves profile images out of the model details it keeps for chats, so Redis no longer holds them and the list loads faster. Functions and Tools now get the model without its profile image.
Measured with 50 custom models, one server process. Each number is the middle value of 5 runs of 60 list loads:
| | before | after |
|---|---|---|
| Redis, 4 images (20 MB as stored) | 64 ms | 45 ms |
| Redis, 12 images (57 MB as stored) | 150 ms | 86 ms |
| No Redis, 4 images | 47 ms | 45 ms |
| Redis memory used by the model list, 4 / 12 images | 24 / 71 MB | 35 KB |
Fixes#34594
Opening a long Markdown file from a terminal's Files froze the whole tab, and typing inside a code block froze it again on every keystroke. The editor guessed the language of every code block that names none by trying every language it knows, about 15 ms per block, so a 750 KB file with a thousand such blocks blocked the page for 19 seconds. Code blocks with no language, or one the editor does not know, now show as plain text, as chat and the read-only file view already show them, and the same file now freezes the page for about 2 seconds. Code blocks that name a language the editor knows are highlighted as before. The same goes for the other places that use this editor, such as notes and the message box with rich text on.
Fixes#34486
When the sign-in provider was unreachable, rate limited or answered with a 5xx while Open WebUI refreshed the user's token, the stored login session was deleted. Tools and connections set to use the user's sign-in token then got no token until the user signed out and back in. Now such failures keep the session and the next tool call refreshes again, while a refresh the provider rejects still removes it, matching the earlier fix for MCP connections.
Fixes#34578
Signing in past OAUTH_MAX_SESSIONS_PER_USER prunes a user's oldest stored OAuth sessions, a back-channel logout from the identity provider ends all of them, and connecting an MCP server again replaces the old session for that server. Pruning and the reconnect logged nothing, and the provider logout logged one info line per user without saying which sessions it removed, so an admin could not trace why a user's provider login was gone. Each of these deletions now logs a warning with the user, the session ID and the reason, the same way a failed token refresh already does.
Fixes#34580
Signing out left the browser's stored OAuth session in place, and once a person reached OAUTH_MAX_SESSIONS_PER_USER the next sign-in removed the sessions created first. A device signed in early and still in use lost its session to ones from browsers that had already signed out. Signing out now removes that browser's stored session when the person still has another one for the same provider, and the limit removes the sessions that were signed in or refreshed least recently. A person's only session stays after sign-out, because automations and API key calls keep using it.
Fixes#34581