Release v0.5.1

Neat Apps v0.5.1

Neat Apps v0.5.1 moves the application catalog to Neat Library v0.5.0. It adds RF-DETR detection and segmentation and FastFlow anomaly detection, makes MLA-accelerated text-to-speech the default in Neat GenAI Studio, and adds an OpenAI-compatible audio API to the Studio.

Highlights

RF-DETR detection and segmentation

A new application runs RF-DETR Small and Medium object detection and RF-DETR Medium instance segmentation on live H.264, H.265, and MJPEG RTSP streams, in C++ and Python (#470). The task and model size are selected in the configuration file, and Insight receives the video with detection boxes or segmentation polygons. A one-frame queue drops stale frames when inference falls behind, so the output stays live.

Input resolution	Codec	Small detector	Medium detector	Medium segmenter
720p	H.264, H.265, MJPEG	Up to 70 FPS	Up to 50 FPS	Up to 40 FPS
1080p	H.264, H.265, MJPEG	Up to 70 FPS	Up to 50 FPS	Up to 40 FPS
4K	H.264	Up to 60 FPS	Up to 50 FPS	Up to 40 FPS

Measured end to end on Modalix under sustained load, with input streams faster than the application can process; each figure is the maximum observed inference output rate. Segmentation uses mask_grid_size: 108.

FastFlow anomaly detection

A new application runs a FastFlow anomaly model on the MLA and shows defects as a heatmap over the live video in Insight, marking each frame OK or ANOMALY (#532, #535). FastFlow learns from images of good parts only. The default model comes from Model Zoo 2.1.3, and a sample inspection clip is included.

Neat GenAI Studio audio

• MLA text-to-speech: Supertonic 3 is the default speech engine, with ten selectable voices (#492). On a Modalix DevKit running Neat 0.4.0 alongside Whisper and LFM2.5-230M, it generated a 33-second utterance in 1.5 seconds.
• OpenAI-compatible audio API: the Studio serves /v1/audio/speech, /v1/audio/transcriptions, /v1/audio/translations, and /v1/audio/voices, so existing OpenAI clients can use on-device speech (#473).
• Audio playground at /playground/ for trying speech, transcription, translation, and a hands-free voice loop.
• Backend-only mode (./run.sh --backend-only) runs the APIs without the web UI, for Insight and other front ends.

Fixes

• YOLOv8 instance segmentation reports class probabilities correctly, so decode.score_threshold filters as documented (#504).
• Eight streaming applications stop when their source ends and report runtime errors, instead of waiting forever or reporting success (#554).
• SuperPoint downloads its model from the published Model Zoo (#469).
• Application README workflows were reproduced on a Modalix DevKit and corrected where they could not be followed (#523).

Installation

sima-cli neat install apps@v0.5.1

The installer also installs the Neat Library version this release was built against.

Compatibility

Component	Version
SiMa.ai Platform	2.1.3
Neat Library	v0.5.0 (7274b4eeb797)
Neat Insight	Latest release (v0.0.7 at release time)

Apps 0.5.1 uses Model Zoo 2.1.3 for documented model downloads.

Migration notes

• Apps v0.5.1 requires Neat Library v0.5.0. Rebuild native applications built from source against libsima_neat.so.5.
• Rerun Neat GenAI Studio's setup.sh to install Supertonic and the updated dependencies (Flask 3.1 or newer).

Contributors

@Onat-Inak, @Yukkta-Seelam, @michaelschlosserx, @dotimothy

Full Changelog: v0.5.0...v0.5.1