LocalAI v4.12.0 Released: Web UI Refresh and Streaming Audio

LocalAI v4.12.0 Released: Web UI Refresh and Streaming Audio

At a Glance

Item Value
Repository mudler/LocalAI
Version v4.12.0
Published 2026-10-22
License MIT
Source type Primary source (the publisher itself)

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

Overview

mudler/LocalAI is an open-source AI engine (written in Go under the MIT license) that allows users to run various models locally—including LLMs, vision, audio, image, and video—regardless of hardware.

Version v4.12.0 brings a complete overhaul of the Web UI, streaming audio responses, native decision routing, and advanced control features for model memory and prompt states. With numerous changes to speech synthesis, model behavior, and privacy protection, operational flexibility in local environments is significantly improved, and users utilizing these features are recommended to update.

Breaking Changes and Deprecations

Several changes have been made to chat-related endpoints and parameters in this release. The dedicated pickup function for modifying inference load during chats, specifically the individual /reasoning action, has been removed.

Before After
Controlled via /reasoning action Removed (Integrated into the UI’s reasoning_effort picker)

This change affects users who previously used fine-grained control over reasoning effort in the chat interface or called /reasoning directly via the API. Moving forward, please use the UI on the chat screen or the integrated parameters.

Key Changes

Web UI Refresh and Improved Model Management

The Web UI has been redesigned using a common CSS/SVG kit and an ink/teal theme, making models and studio workflows much clearer. Remote manual imports are now executed as background gallery jobs, allowing users to track progress, failures, and cancellations.

Model storage aggregation now distinguishes between shared files and model-specific files. Please note that the storage reclaimed upon deletion may not always match the displayed size.

Streaming Voice and Enhanced Audio Analysis

For text-to-speech engines like Qwen TTS, a live streaming feature has been added that feeds LLM fragments sequentially to start audio generation early. This enables audio output to begin without waiting for the LLM to complete its full response.

Parakeet audio bundles can now contain multiple roles such as automatic speech recognition (ASR), VAD, and speaker diarization, while also supporting speaker verification via voice_verify_threshold. This impacts users building advanced audio-related pipelines.

Notice Regarding Proprietary Weights

YuE2, newly added to the audio generation path, is subject to the CC-BY-NC-4.0 license and is restricted to non-commercial use. Users expecting commercial deployment should take note, as it cannot be adopted in such environments.

Model Memory and Prompt State Control

Additions include moe_cache_mib for caching MoE model experts on the host side, lazy_mode to enable lazy loading, and slot saving and restoration functions via /v1/slots/save and /v1/slots/restore. These allow for memory usage optimization according to workloads and the persistence of prompt states.

Slot persistence functions operate only on POSIX environments, retaining up to 32 save states per model. Saved files are treated as sensitive data, requiring appropriate access management.

Supported Models and Hardware

This version expands supported models and hardware across a wide range of fields, including audio/music generation, image processing, and motion capture.

Audio and Music Generation Models

Qwen TTS has introduced the TTSLive backend RPC, enabling it to output audio while sequentially receiving generated text from an LLM. This allows for real-time interaction where audio playback starts without waiting for the entire text generation to finish.

The Parakeet model bundling feature now handles multiple roles collectively, including automatic speech recognition (ASR), voice activity detection (VAD), speaker diarization, acoustic event detection for environmental sounds, and a speaker encoder. It also supports voice embeddings using the speaker encoder and matching via cosine distance.

In the music generation domain, YuE2 is now supported in the audio.cpp generation path. However, note that YuE2 model weights are governed by the CC-BY-NC-4.0 license, which prohibits commercial use.

Furthermore, OmniVoice now allows generation time settings (up to 30 seconds) and post-processing parameter controls, while Pocket TTS supports language settings during model loading (e.g., options: ["language:french"]).

In the model gallery, three streaming assistant presets have been added: qwen3-streaming-assistant, vits-streaming-assistant, and piper-streaming-assistant. Because these provide only pipeline configuration settings, users must separately prepare and install the corresponding LLM, transcription, VAD, and TTS models to run them.

Image/Video Processing and Motion Capture Models

For image processing, a native mode for ESRGAN-based upscaling has been introduced to perform tiling without retaining diffusion model context. This enables memory-efficient high-resolution scaling.

In video and multimedia, MiniMax-H3 now allows reference images, reference videos, and WAV audio to be set as input fields (available via API only in this version).

Additionally, real-time motion capture using GEM-X via the new gemxcpp backend has been introduced. Pose data in SOMA-77 or SMPL-24 formats can be extracted in real-time from streamed RGB video frames. The hardware supports CPU and Vulkan, running on CPU on macOS. Note that the use of GEM-X weights is subject to NVIDIA’s model license terms.

Vector Databases and Execution Launcher

For persistent vector management, the qdrant-store backend has been added to utilize Qdrant via /stores/* endpoints. It supports TLS communication, custom certificate authorities, API key configuration via environment variables, and batch write control.

Additionally, a lightweight AppImage launcher for Linux x86_64 has been added to the distribution packages. Running this launcher requires glibc 2.35 or higher and appropriate host graphics libraries installed on the system.

How to Get It

Specific commands for updating to LocalAI v4.12.0 are not provided in the distribution materials. Please refer to the official LocalAI v4.12.0 Release Page to check update procedures tailored to your deployment format, whether Docker, compilation, or various binary distributions.

If you are deploying the newly provided AppImage launcher in a Linux environment, please ensure beforehand that your environment meets the operational requirements of glibc 2.35 or higher along with the necessary host graphics libraries.

Related Articles

What to Read Next

Sources