Open WebUI 0.12: Realtime Voice, Skills and Shared Chats

Last updated: November 2, 2026

Open WebUI 0.12, released in six versions from October 10 to October 29, 2026, adds Realtime voice calls through OpenAI or Azure OpenAI, multi-file skills with version history, shared chats that others can reply in, per-model controls such as reasoning effort, required two-factor sign-in, encrypted channel chats and an admin session explorer.

Contents

What’s new in Open WebUI 0.12
Realtime voice calls
Skills, compaction and context
Shared chats, channels and apps
Security and admin controls
How to update
Practical workflows
Known limitations
Frequently asked questions

Open WebUI is a self-hosted chat interface that connects to local runtimes such as Ollama and to OpenAI-compatible APIs. The 0.12 series is one of its largest feature cycles. Everything below comes from the project’s official GitHub release notes for v0.12.0 through v0.12.5 and its documentation. If you run models locally, our Ollama new features guide and LM Studio Bionic guide cover the runtimes underneath.

What’s new in Open WebUI 0.12

Version	Date (UTC)	Headline changes
0.12.0	Oct 10	Realtime voice calls, VRM avatars, two-factor sign-in, replies in shared chats, model controls, multi-file skills, knowledge base file browser, redesigned model editor
0.12.1	Oct 14	External apps in channels, Compare revision view and /compare, .skill export, Azure OpenAI Realtime, input and output token columns in analytics
0.12.2	Oct 14	MinerU 4.x support, {{MODEL_ID}} in custom headers, call_tool for Tool Search, several security fixes
0.12.3	Oct 16	End-to-end encrypted channel chats, workspace search, lighter Redis streaming, skill read limits
0.12.4	Oct 25	Skills loaded on demand, visible compaction markers, admin control over which UI controls users see
0.12.5	Oct 29	Analytics session explorer, pages in notes, admin API key rotation, faster knowledge base and chat search

Realtime voice calls

Before 0.12, voice mode chained speech-to-text, the chat model and text-to-speech. The new Realtime call mode talks through an OpenAI Realtime voice model, by default gpt-realtime-2.1-mini with the Marin voice, which handles small talk itself and hands real questions to the chat’s selected model with its history and tools. The full reply still appears in the chat, according to the v0.12.0 release notes.

Open Admin Settings, go to Audio, and in Voice calls switch Call mode from Standard to Realtime (or set AUDIO_REALTIME_ENABLED).
Enter the API base URL and key, or set AUDIO_REALTIME_OPENAI_API_BASE_URL and AUDIO_REALTIME_OPENAI_API_KEY.
To use Azure instead (0.12.1 and later), pick Azure OpenAI as the Provider and enter the endpoint and deployment name, signing in with an API key or Microsoft Entra ID.
Grant users the Allow Call permission.
Optionally set a per-model Realtime Voice, or upload a rigged VRM avatar of up to 25 MiB with up to 16 named gestures in the model editor.

Tool approvals and questions must still be answered in the chat, and calls end after an hour. For background on OpenAI’s voice lineup, see our GPT-Live voices guide.

Skills, compaction and context

Workspace skills can now hold scripts, references and templates next to SKILL.md, and every save keeps a version you can compare or restore. Skills import from a ZIP, JSON or a single SKILL.md, and 0.12.1 added a single-file .skill export. Models can create and edit skills themselves, so /skills:create saves straight to the workspace without a terminal.

Two later settings help with context cost. Version 0.12.3 added Skill Read Limit and Skill Preview Length (SKILL_CONTENT_MAX_CHARS and SKILL_PREVIEW_MAX_CHARS, defaulting to 20000 and 5000 characters). Version 0.12.4 added Load selected skills on demand (ENABLE_SKILL_LAZY_LOADING), which lists selected skills for the model to load when needed instead of sending full instructions with every message. It is off by default and needs the built-in Skills tool.

Context compaction also improved. With native function calling, 0.12.0 compacts between tool rounds in long tool runs. In 0.12.4, a “Context compacted” line marks where older messages were summarized, and compaction now runs in temporary chats and short chats already over the threshold.

Model controls, also new in 0.12.0, let anyone who can edit a model add named options, such as a Thinking control whose Low, Medium and High options set reasoning effort. Users pick an option from a knobs button next to the model selector.

Shared chats, channels and apps

Sharing a chat or folder now has a Sharing mode that can be switched from Clone only to Allow replies, so recipients keep writing in the same conversation with live streaming and typing indicators. From 0.12.3, people allowed to edit can branch and regenerate any reply, and the top bar shows who else has the chat open.

Channels gained external apps in 0.12.1: administrators enable Apps (ENABLE_APPS), connect an app by its HTTPS address and review its permissions, and channel members can @mention it for replies, cards, buttons or file downloads.

Security and admin controls

Feature	Version	How to turn it on
Required two-factor sign-in	0.12.0	Admin Settings > Authentication, or ENABLE_MFA
End-to-end encrypted channel chats (MLS, messages up to 16 KiB, not stored)	0.12.3	Admin Settings > General > Encrypted channels, or ENABLE_ENCRYPTED_CHANNELS, plus the Encrypted Channels permission
Hide individual controls from users	0.12.4	Admin Settings > Interface > Visible controls, or UI_VISIBILITY
Analytics session explorer	0.12.5	Sessions tab in Admin Analytics; requires ENABLE_ADMIN_CHAT_ACCESS
Rotate or delete a user’s API key	0.12.5	Edit User dialog

Encrypted chats need websocket connections with websocket server logging off, and HTTPS in production.

How to update

The official updating guide gives this manual Docker sequence. Chats, users and settings live in the open-webui volume, which is preserved.

Stop and remove the container: docker rm -f open-webui.
Pull the image: docker pull ghcr.io/open-webui/open-webui:main, or replace :main with a pinned version.
Recreate it with the same -v open-webui:/app/backend/data volume and the same WEBUI_SECRET_KEY.
With Docker Compose, run docker compose pull and then docker compose up -d.

The guide warns that multi-instance deployments should stop all replicas and let one instance run the database migration before the rest return. The quick start notes that without a fixed WEBUI_SECRET_KEY, every recreated container logs everyone out.

Practical workflows

Cheaper long sessions. Turn on skills on demand and keep compaction enabled so large skill libraries and tool output do not fill the context window.
Per-model routing. Use {{MODEL_ID}} in connection or tool server headers (0.12.2) so a gateway can log or route usage per model.
Revising prompts. Ask a model for a rewrite, then use /compare to step through each change and apply only the ones you want.

If you also work in terminal agents, compare the skills approach with our Claude Code guide.

Known limitations

Realtime calls depend on an OpenAI or Azure OpenAI Realtime endpoint and end after one hour.
Model controls cannot be set on pipe, direct connection or arena models.
Encrypted chats are temporary: they clear when you leave or reload and end when channel membership changes.

Frequently asked questions

What is new in Open WebUI 0.12?

Open WebUI 0.12.0 to 0.12.5 shipped between October 10 and October 29, 2026. The main additions are Realtime voice calls, multi-file skills with version history, replying in shared chats, model controls, two-factor sign-in, end-to-end encrypted chats in channels and an analytics session explorer.

How do I turn on Realtime voice calls in Open WebUI?

An administrator switches Call mode from Standard to Realtime in the Voice calls section of Audio settings, or sets AUDIO_REALTIME_ENABLED. The default voice model is gpt-realtime-2.1-mini with the Marin voice, and users need the Allow Call permission. Calls end after one hour.

Can Open WebUI require two-factor authentication?

Yes. Since 0.12.0 an administrator can require an authenticator app for every user under Admin Settings > Authentication or with ENABLE_MFA. Users get ten single-use recovery codes, and the open-webui mfa reset command issues a one-time recovery token.

How do I update Open WebUI to 0.12 with Docker?

Remove the container with docker rm -f open-webui, pull ghcr.io/open-webui/open-webui:main, and run the container again with the same open-webui volume and WEBUI_SECRET_KEY. Your data stays in the volume.

What does Load selected skills on demand do?

Added in 0.12.4, it lists selected skills for the model to load only when needed instead of sending their full instructions with every message. It is off by default, needs the built-in Skills tool and is set under Admin Settings > General or with ENABLE_SKILL_LAZY_LOADING.

Sources

Image: Open WebUI logo from the Open WebUI GitHub organization, public domain, via Wikimedia Commons. Placed unchanged on a white background. FreeAnswers is not affiliated with Open WebUI.

Open WebUI 0.12: Power-User Guide to the October 2026 Releases