generated: '2026-08-27'
method: searched
source: https://github.com/mudler/LocalAI/releases
docs: https://github.com/mudler/LocalAI/releases
blog: https://localai.io/blog/
scheme: semver
current_version: v4.9.0
current_version_released: '2026-08-20'
window: v4.5.6 (2026-06-30) through v4.9.0 (2026-08-20), the ten most recent releases at probe time
format_note: >-
  Two channels carry the same events at different resolutions. GitHub Releases is the machine-readable
  one — every tag dated, minor releases carrying a hand-written narrative and patch releases carrying
  auto-generated "What's Changed" PR lists produced from .github/release.yml. localai.io/blog/ carries a
  long-form "What landed in LocalAI <minor>" post per minor release. Neither is a structured, per-endpoint
  API changelog: an agent cannot learn from either which operations changed shape in a given release.
entries:
- version: v4.9.0
  date: '2026-08-20'
  type: minor
  breaking: true
  highlights:
  - >-
    Authentication is now DENY-BY-DEFAULT — every HTTP route requires credentials unless it appears on the
    published anonymous discovery/bootstrap list. This is a behavioural break for any deployment that
    relied on unauthenticated reads.
  - Chat gained end-to-end context compression.
  - Models and backends each got one canonical page in the UI instead of three.
  - vllm-cpp grew a video modality serving MiniMax-H3 with a real audio track.
  scope: 13 days, 146 pull requests
- version: v4.8.2
  date: '2026-08-07'
  type: patch
  breaking: false
  highlights:
  - Gallery falls back to mirrors and a cached index when the primary source fails.
  - Dependency bumps (actions/stale, actions/checkout) and model-gallery checksum updates.
- version: v4.8.1
  date: '2026-08-06'
  type: patch
  breaking: false
  highlights:
  - Contain malformed GGUF metadata in VRAM estimation.
  - Documentation and vendored-engine checksum updates.
- version: v4.8.0
  date: '2026-08-05'
  type: minor
  breaking: false
  highlights:
  - >-
    vllm.cpp introduced — a C++20 engine maintained by the LocalAI team, begun as a vLLM port and now
    carrying its own featureset, shipping as the vllm-cpp backend in alpha development builds.
  - 3D generation added as a new modality.
  - A multi-family audio.cpp engine.
  - Gallery entries now install the build the host hardware can actually run.
  - A reliability pass on distributed mode.
  scope: 22 days, 386 pull requests, three new modalities
- version: v4.7.1
  date: '2026-07-14'
  type: patch
  breaking: false
  highlights:
  - Only inject llama.cpp serving options on the llama.cpp path.
- version: v4.7.0
  date: '2026-07-14'
  type: minor
  breaking: false
  highlights:
  - UI-managed voice cloning library with consented recording or upload.
  - Local video and audio-driven avatar generation.
  - Interleaved reasoning that travels with tool calls.
  - New audio engines — one-pass diarized transcription, F5-TTS, and true streaming TTS.
  - Reliability fixes across auth, transcription and the model gallery.
- version: v4.6.2
  date: '2026-07-06'
  type: patch
  breaking: false
- version: v4.6.1
  date: '2026-07-06'
  type: patch
  breaking: false
- version: v4.6.0
  date: '2026-07-04'
  type: minor
  breaking: false
- version: v4.5.6
  date: '2026-06-30'
  type: patch
  breaking: false
narrative_posts:
- title: What landed in LocalAI 4.8
  url: https://localai.io/blog/what-landed-in-localai-4-8/
- title: What landed in LocalAI 4.3
  url: https://localai.io/blog/what-landed-in-localai-4-3/
- title: What landed in LocalAI 4.2
  url: https://localai.io/blog/what-landed-in-localai-4-2/
- title: What landed in LocalAI 4.1
  url: https://localai.io/blog/what-landed-in-localai-4-1/
- title: What landed in LocalAI 4.0
  url: https://localai.io/blog/what-landed-in-localai-4-0/
- title: What landed in LocalAI 3.10
  url: https://localai.io/blog/what-landed-in-localai-3-10/