Open WebUI v0.12.0 Released: Real-Time Voice & VRM Avatars

Open WebUI v0.12.0 Released: Real-Time Voice & VRM Avatars

At a Glance

Item Value
Repository open-webui/open-webui
Version v0.12.0
Published 2026-10-11
License NOASSERTION
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

The latest version of “Open WebUI" (v0.12.0), a web interface that allows intuitive operation of various models such as local LLMs and the OpenAI API, has been released. Open WebUI is built in Python and serves as a frontend environment that bundles local environments like Ollama and external APIs to provide chat and agent functions.

The biggest changes in this update are the inclusion of a “real-time voice call function" integrated with OpenAI’s Realtime audio models and “VRM 3D avatar display." This allows users to experience an advanced voice agent locally, where a low-latency real-time voice model handles casual small talk while complex reasoning and tool execution are automatically delegated to the model selected on the chat side.

Key Changes

Addition of Real-Time Voice Calls and VRM Avatar Display

“Realtime" has been added to the call mode in the voice settings, enabling two-way voice calls using OpenAI Realtime voice models (defaulting to the Marin voice gpt-realtime-2.1-mini). The voice model itself directly replies to casual chats, while full-scale tasks and questions are passed along with conversation history and tools to the currently selected model for processing. Calls automatically end after a maximum of 1 hour, and usage requires the “Allow Call" permission. Furthermore, model settings now allow registering rigged VRM avatars up to 25 MiB, VRMA clips up to 10 MiB corresponding to idle, listening, and speaking states, and up to 16 types of gestures. This makes it possible to display a 3D character that responds with rich facial expressions while lip-syncing to the voice output.

Collaborative Replies and Forking in Shared Chats

In addition to the conventional clone-only option, “Allow replies" has been added to the sharing mode when sharing chats or folders with other users or groups. Enabling this setting allows sharing partners to collaborate by posting messages directly on the same chat without duplicating the conversation. Sender names and icons are displayed for messages, which are streamed in real time along with typing indicators. Each user can edit or regenerate only their own messages. Additionally, a “fork" function has been added, allowing users to replicate conversations up to that point into their own environment from completed replies in shared links or read-only chats.

Introduction of Model Controls

It is now possible to create “Model Controls" in the model edit screen to define unique operational parameters for each model. For example, controls for switching the thinking level (Low, Medium, High) of a reasoning model can be placed in menu or slider formats. Authorized users can easily switch parameters via the knob button in the input field, and the selected content is remembered for each model. Note that this is not applied to pipe models, direct connection models, arena models, or when executing automated tasks.

Tool Search and Context Compaction for Large-Scale Tool Operations

As an optimization for environments connected to numerous tools or MCP servers, the experimental feature “Tool Search" has been added. This mechanism removes long tool definitions from the initial prompt, allowing models to dynamically retrieve definitions via the search_tools tool as needed. By specifying the character count threshold (Deferral Threshold) and always-loaded tools (Always Loaded Tools), it becomes easier to maintain prompt caching. Additionally, when executing numerous tools consecutively with native function calling, processing that intersperses context compaction between turns has been implemented, preventing models from exceeding context limits even during lengthy tool executions.

Multi-Factor Authentication (MFA) and Enhanced Security

Two-factor authentication using authenticator apps can now be made mandatory for all users via administrator settings and the ENABLE_MFA environment variable. Setup is performed via a QR code upon the next login, issuing 10 single-use recovery codes. For users who lose their authenticator app, administrators can issue a one-time recovery token by running the CLI command `open-webui mfa reset`.

Major Speedup of JSON Processing and Performance Improvements

orjson has been adopted as the default library for JSON read/write processing across the entire application. As a result, encoding speed for request/response processing, streaming reception from providers, and WebSocket live updates has improved by about 17x, and decoding speed by about 3x. In addition, hybrid searches targeting Milvus 2.5 and later have transitioned to native processing that directly utilizes Milvus’s built-in BM25 full-text search. This eliminates the traditional load of loading all chunks into the WebUI side for scoring, significantly reducing memory consumption and increasing speed.

Propagation of Finish Reason to Outlet Filters

For plugin developers, finish_reason is now attached to assistant messages upon response completion in OpenAI-compatible models and Ollama models. This allows outlet filters to accurately determine whether a model finished naturally, reached the token limit, or stopped due to a tool call, without having to monitor stream filters chunk by chunk. At the same time, the behavior where the reported value was always stop when disconnected due to the token limit while using Ollama has been fixed, and length is now reported appropriately.

Specifications

Integrated Models and External Services

For the real-time voice call function, newly added support includes OpenAI’s Realtime voice models (defaulting to the Marin voice gpt-realtime-2.1-mini). While the voice model itself handles small talk, full-scale tasks and tool execution can be handed off to various models selected on Open WebUI for cooperative operation.

Additionally, in OpenAI-compatible models and various local models running via Ollama, the finish reason upon model call completion is now accurately passed to outlet filters. Even when a Ollama model hits the token limit and disconnects, length is appropriately reported instead of the traditional stop.

As an external vector database, versions Milvus 2.5 and later are now supported. Native hybrid search utilizing the built-in BM25 full-text search is available, eliminating the need to fetch all chunks to the Open WebUI side for recalculation as was done previously.

Avatar Formats and Document Generation

Model settings now support uploading rigged VRM format 3D avatars (up to 25 MiB). Alongside this, VRMA motion clips corresponding to idle, listening, and speaking states (up to 10 MiB) and up to 16 unique gestures can be registered.

Furthermore, in the code interpreter environment, using the bundled libraries makes it possible to directly generate Microsoft Word documents and PowerPoint slides and open them in Office, even in offline environments without an internet connection.

Execution Environment and Accelerators

For the execution environment, pip installation in a Python 3.11 environment is supported, as well as Docker container environments. As Docker images, standard CPU-friendly images, GPU-accelerated images using the Nvidia CUDA toolkit (:cuda tag), and images with Ollama bundled inside the container (:ollama tag) are available.

How to Get It

Update using pip or Docker, depending on your Open WebUI deployment environment.

Update Procedure Using pip

If you are running directly in a Python environment, verify that Python 3.11 is being used, run the following command to update the package to the latest version, and restart the server:

pip install open-webui

After the update is complete, start the service as usual:

open-webui serve

Once started, access http://localhost:8080 from your browser to check its operation.

Update Procedure Using Docker

If you are running in a Docker container, stop and remove the existing container, then pull the latest image and recreate it. To prevent the loss of databases and various settings, make sure to specify a volume mount (-v open-webui:/app/backend/data).

For a standard configuration connecting to Ollama running on the host device, run the following command:

docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:main

To start with CUDA acceleration enabled via Nvidia GPU, run the following command:

docker run -d -p 3000:8080 --gpus all --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:cuda

To use the integrated image with Ollama bundled inside the container in a GPU environment, run the following command:

docker run -d -p 3000:8080 --gpus=all -v ollama:/root/.ollama -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:ollama

After starting the container, you can begin using it by accessing http://localhost:3000 from your browser.

Related Articles

What to Read Next

Sources