v4.4.0
📦 localaiView on GitHub →
✨ 11 features🐛 2 fixes🔧 8 symbols
Summary
LocalAI 4.4.0 is a major multimodal and distributed release, introducing two new audio backends (Parakeet and CrispASR), native object detection, video understanding/generation, and a new intelligent middleware layer for routing and PII filtering.
Migration Steps
- If using distributed mode and relying on the previous round-robin behavior exclusively, disable prefix-cache-aware routing by starting with `--distributed-prefix-cache=false`.
- If using NATS in distributed mode, configure JWT authentication and TLS/mTLS, as unauthenticated access is now restricted.
- If transferring large files in distributed mode, be aware that transfers now use Content-Range semantics for resumability.
✨ New Features
- Introduced two new audio backends: parakeet.cpp (NVIDIA NeMo Parakeet ASR) and CrispASR (ASR and TTS engine).
- Added native object detection and segmentation support via the rfdetr-cpp backend.
- Enabled video understanding capabilities in llama-cpp via mtmd.
- Integrated LTX-2 video generation support in the stablediffusion-ggml backend.
- Distributed mode now features prefix-cache-aware routing enabled by default.
- Distributed mode supports resumable HTTP file transfers for large models.
- Implemented a new intelligent middleware layer for capability-based model routing, PII filtering/redaction, and cloud-model proxying.
- Added an interactive `local-ai chat` CLI tool with commands like /models, /model, and /clear.
- RAG agents now provide source citations appended as a clickable 'Sources:' block.
- Added TTS per-request control via `instructions` and a generic `params` map (supporting Qwen3-TTS VoiceDesign / CustomVoice, Chatterbox).
- Support for Gemma 4 QAT family, QAT-matched MTP speculative-decoding bundles, Ideogram4, and LTX-2.3 22B GGUFs.
🐛 Bug Fixes
- Fixed a long run of reasoning and tool-call streaming issues.
- Implemented security hardening in pkg/httpclient to close a credential-leak class across every outbound HTTP client (GHSA-3mj3-57v2-4636).