# LocalAI v3.5.0 — Release & Upgrade Intelligence
- **Product:** LocalAI (https://localai.io)
- **Version:** v3.5.0
- **Released:** Sep 03, 2025
- **Impact Score:** 5/10
- **Breaking Changes:** Not verified
- **Canonical Analysis:** https://changelogs.us/localai/updates/v3.5.0/

## Executive Summary (TL;DR)
LocalAI v3.5.0. Source highlight: Welcome to LocalAI 3.5.0! This release focuses on expanding backend support, improving usability, refining the overall experience, and keeping reducing footprint of LocalAI, to make it a truly portable, privacy-focused AI stack. We’ve added several new backends, enhanced the WebUI with new features,

## New Features
- 🖼️ **Expanded Backend Support:** Welcome to MLX!  **mlx**, **mlx-audio**, **mlx-vlm** are now all available in LocalAI. We also added support to WAN for video generation, and a CPU and MPS version of the diffusers backend! Now you can generate and edit images from MacOS or if you don't have any GPU (albeit slow).
- Access to the WebUI and Application Folder
- **System Backend:** Added a new "system" backend path (`LOCALAI_BACKENDS_SYSTEM_PATH` or via command-line arguments) defaulting to `/usr/share/localai/backends`. This allows specifying a read-only directory for backends, useful for package management and system-wide installations.
- **Darwin CI Builds:** Added support for building some Go-based backends (Stablediffusion and Whisper) on Darwin (macOS) in the CI pipeline.
- chore(build): Convert stablediffusion-ggml backend to Purego
- feat(backends): add system backend, refactor

## Improvements
- ✨ **WebUI Enhancements:** Download model configurations, a manual model refresh button, streamlined error streaming during SSE events, and a stop button for running backends. Models now can also be imported and edited via the WebUI.
- 🚀 **Performance & Architecture:** Whisper backend has been rewritten in Purego with integrated Voice Activity Detection (VAD) for improved efficiency and stability. Stablediffusion also benefits from the Purego conversion.
- **Enhanced support for MacOS**: whisper, diffusers, llama.cpp, MLX (VLM, Audio, LLM), stable-diffusion.cpp will now work on MacOS!
- **Download Model Config:** A "Get Config" button in the model gallery lets you download a model’s configuration file without installing the full model. This is perfect for custom setups and easier integration.
- **Manual Model Refresh:** A new button allows you to manually refresh the on-disk YAML configuration, ensuring the WebUI always has the latest model information.
- **Backend Stop Button:** Quickly stop running backends directly from the WebUI.

## Bug Fixes
- ✅ **Bug Fixes & Stability:** Resolutions to AMD RX 9060XT ROCm errors, libomp linking issues, model loading problems on macOS, CUDA device detection improvements, and more.
- **Streamlined Error Handling:** Errors during SSE streaming events are now displayed directly to the user, providing better visibility and debugging information.
- **AMD RX 9060XT ROCm Error:** Fixed an issue causing errors with AMD RX 9060XT GPUs when using ROCm. This error, "ROCm error: invalid device function", occurred because of device function incompatibility.  The fix involves updating the ROCm image and ensuring the correct GPU targets are specified during compilation.  Recommended kernel versions and verification steps for GPU detection are available here.
- **libomp Linking:** Resolved a missing `libomp.so` issue on macOS Docker containers.
- **macOS Model Loading:** Addressed a problem where models could not be loaded on macOS. This was resolved by bundling necessary `libutf8` libraries.
- **CUDA Device Detection:** Improved detection of available GPU resources.