v4.3.2
📦 localaiView on GitHub →
✨ 5 features🐛 15 fixes🔧 1 symbols
Summary
This release introduces significant new features like model routing, PII filtering, and object detection backends, alongside numerous dependency updates and bug fixes across distributed systems and model handling.
✨ New Features
- Implemented model routing, PII filtering, and Cloud model proxies via middleware.
- Added gated X-LocalAI-Node response header (middleware + wrapper) for distributed systems.
- Introduced LTX-2 support and LTX-2.3 GGUF gallery entries for stablediffusion-ggml.
- Updated swagger documentation.
- Added rfdetr-cpp native object detection and segmentation backend.
🐛 Bug Fixes
- Fixed VRAM detection for Intel hardware.
- Ensured per-model load information is persisted so the reconciler survives frontend restarts in distributed setups.
- Synchronized gallery OpCache and caches across frontend replicas in distributed setups.
- Added missing diffusion_model flag to all LTX-2.3 gallery variants.
- Set vae_decode_only:false for i2v / flf2v in LTX-2.3 gallery fixes.
- Prevented <think> tags from leaking into content when autoparser is in pure-content mode.
- Fixed muxing of LTX-2 audio into output MP4 for stablediffusion-ggml.
- Stopped healing-marker stubs from gating off content during streaming/tools processing.
- Prevented prefill-misclassified content from being leaked as a trailing reasoning chunk during streaming/tools processing.
- Excluded local-only artifacts from the build context in .dockerignore.
- Enabled timestamp output for qwen-asr when forced_aligner is configured.
- Fixed Nemo ASR models to extract Hypothesis.text for TDT/RNNT models.
- Shared a single /api/operations poller across consumers in the React UI.
- Polished the 'Fits in my GPU' filter in the React UI to use the design-system Toggle.
- Forced .check() on hidden Toggle input in the fits-filter e2e test for the React UI.