Small engines.
Real capability.

Cache-aware Parakeet speech recognition in C++ and GGML.

Long-form transcription, speakers, and timestamps in one pass.

Native sound-event recognition across 527 AudioSet classes.

Voxtral text-to-speech implemented in pure C.

Native voice cloning, synthesis, transcription, and diarization.

Echo cancellation, noise suppression, and dereverberation.

Speaker recognition and voice analysis without Python inference.

Detection, recognition, demographics, and anti-spoofing.

Native object detection and instance segmentation.

Open-vocabulary object detection and visual grounding.

Monocular metric depth and camera-pose estimation.

Turn ordinary photos into 3D Gaussians without camera poses.

Standalone GGML PII and named-entity detection.

A local-first vector database shipped inside LocalAI.

MoE-aware GGUF recipes that preserve quality through sparsity.