●Change8

v3.9.0

Breaking Changes
📦 localaiView on GitHub →
⚠ 2 breaking✨ 7 features🐛 2 fixes🔧 4 symbols

Summary

LocalAI 3.9.0 introduces significant stability and resource management features, including an Agent Jobs panel for scheduling and an automatic Smart Memory Reclaimer with LRU model eviction. This release also drops x86_64 Mac support and updates data storage paths.

⚠️ Breaking Changes

  • Data storage path has been updated from /usr/share to /var/lib to follow Linux conventions for mutable data. Users relying on the old path for configuration or data access must update their paths.
  • x86_64 Mac support has been dropped and is no longer maintained in GitHub Actions. Users on x86_64 Macs should migrate to ARM64 architecture.

Migration Steps

  1. If you were relying on data stored in the old path (/usr/share), update your configuration or scripts to point to the new path (/var/lib).
  2. Users on x86_64 Macs should plan to migrate to ARM64 architecture as x86_64 support is discontinued.
  3. If using the single_active_backend mode, note that it now defaults to LRU=1 for backward compatibility.

✨ New Features

  • Introduction of the Agent Jobs panel in the web UI and API for scheduling and running agentic tasks using cron syntax or API triggers.
  • Implementation of a Smart Memory Reclaimer that monitors system memory and automatically evicts the least recently used (LRU) models when usage exceeds a threshold to free up GPU/VRAM.
  • Support for LRU (Least Recently Used) model eviction, allowing users to set a maximum number of models to keep in memory, automatically unloading the oldest unused model when the limit is reached.
  • Added support for MLX and CUDA 13 backends.
  • Added support for the VibeVoice TTS backend, including real-time text-to-speech with voice cloning capabilities.
  • StableDiffusion-GGML now supports LoRA for enhanced image generation.
  • UI improvements including fixed navbar ordering, login icon placement, prevention of tool call overflow in chat view, uniformed link paths (e.g., /browse/), and consistent button styling.

🐛 Bug Fixes

  • Fixed an issue where /readyz and /healthz endpoints incorrectly required authentication after v3.8.0, which broke Docker health checks and monitoring tools.
  • Fixed crashes that occurred when importing models from Hugging Face URLs containing subfolders (e.g., huggingface://user/model/GGUF/model.gguf).

🔧 Affected Symbols

/readyz/healthzStableDiffusion-GGMLsingle_active_backend