Release v0.6.0

Neat Apps v0.6.0

Neat Apps v0.6.0 moves the application catalog to SiMa.ai Platform 3.0.0 and Neat Library v0.6.0. It adds applications for inspection, segmentation, multi-stream vision and containers, makes RF-DETR detection about 70% faster in four sizes, scales high-density detection to 60 streams, and turns Neat GenAI Studio into a backend for Insight's GenAI tab.

Highlights

New applications

• PCB defect detection inspects a folder of circuit-board images with a custom YOLO26n model and writes annotated images and per-class defect counts for six defect types. No camera or Insight is required (#474).
• PatchCore anomaly detection learns from good images only and highlights anomalous regions from an image folder, a video file or an RTSP stream. --calibrate builds the memory bank from the compiled model (#475).
• EfficientSAM3 promptable segmentation segments whatever a text prompt names in an H.264 or H.265 RTSP stream and sends mask outlines to Insight. Changing the target is a prompt.text edit (#579).
• Image classification explorer replaces the basic image classifier. It classifies one image or a folder with one or more models (ResNet-50, ResNet-18, EfficientNet-B0, DenseNet-121) and writes an HTML report plus JSON and CSV results (#520).
• Multi-stream BlazePose 3D runs one shared YOLO26 detector and one BlazePose Heavy model across up to four RTSP streams and publishes 2D and 3D landmarks to Insight (#527).
• Multi-stream oriented-box detection runs YOLO26n-OBB across up to four RTSP streams for aerial video and publishes rotated boxes with an upright fallback box (#639).
• Multi-stream multi-model (experimental) runs a different detection, segmentation or pose model on each of up to four RTSP streams (#539).
• Zero-copy MIPI multi-model captures IMX568 video with CPU camera copies disabled by default and runs a selectable YOLO26, SSD-MobileNet V3, ResNet-50 or MiDaS model (#460).
• Parallel apps in Docker builds one container image per application and runs six existing examples side by side on a DevKit (#633).

RF-DETR detection about 70% faster

RF-DETR Small detection reaches 121 FPS and Medium 82 FPS at 720p and 1080p, up from 70 and 50 FPS in Apps 0.5.1: 73% and 64% faster. New Nano and Large sizes reach about 210 and 43 FPS, and Small stays the default (#544, #638).

Input resolution	Codec	Nano detector	Small detector	Medium detector	Large detector
720p	H.264	208	121	82	43
720p	H.265	212	121	82	43
720p	MJPEG	186	121	81	43
1080p	H.264	211	121	82	43
1080p	H.265	205	120	82	43

Measured with the Python application on an otherwise idle Modalix, averaging the inference output rate over at least 20 seconds with 300 FPS input streams. The models are now w8a8_w16a16, and deeper queues between the backbone and transformer let fast models use the MLA fully. H.264 and H.265 video reaches Insight as the original encoded stream.

More streams per device

• High-density detection ships 20-, 30- and 60-stream profiles, replacing 16, 24 and 48. The defaults target 30 FPS at 20 streams, 20 FPS at 30 streams and 10 FPS at 60 streams (#618, #643). Actual capacity depends on the model and stream settings.
• The multi-stream detector, pose estimator and tracker run each application as one pipeline with an encoded tee per stream (#563).
• Multi-stream tracking follows one to five classes, each with its own thresholds and Kalman settings, using a ByteTrack-style tracker (#538).
• The Python tiny-drone tracker's association step fell from 25–67 ms to 3.6–4.8 ms per frame with 30–60 active tracks, keeping up with a 30 FPS stream on Modalix (#581).

Neat GenAI Studio for Insight

• GenAI Studio runs backend-only and serves Insight's GenAI tab (#602).
• Several chat models can stay loaded at once through server.models.max_resident_chat_models; the default remains one (#587).
• Document chat (RAG) works in backend-only mode, and SETUP_RAG=0 skips its dependencies on constrained boards (#602, #622).
• Chat no longer repeats a catalog lookup and a second tokenizer load before generating, which had delayed the first text by about 9 seconds in the reported case. Supertonic speech is batched, and paired DFlash and Gemma MTP packages appear in the catalog (#653).
• setup.sh retries rate-limited and dropped downloads (#631).

Other improvements

• Single-stream object detection and instance segmentation accept a USB camera and pass its MJPEG video to Insight without re-encoding (#644).
• Single-stream instance segmentation handles both YOLO26 and YOLOv8 through model.family, replacing the separate YOLOv8 segmenter (#537).
• The Examples portal has a consistent catalog and detail-page layout with search and category filters (#615, #624).
• BENCHMARK_RESULTS.md is re-measured on Platform 3.0.0 for 59 Model Zoo packages (#662).

Fixes

• Saved H.265 instance-segmentation frames stay paired with their results (#628).
• Multi-stream pose no longer sends a per-frame pose id, so a person's overlay color no longer changes between frames (#625).
• The MIPI camera capture example starts on Platform 3.0.0 by using the camera's default buffer count (#659).
• RF-DETR accepts the batch-one feature shape reported by Neat Library v0.6.0 (#664).
• The PCIe high-density detector exits when the card stops returning results and guards against restarting into a stalled card (#477).
• FastFlow's model-preparation steps pin anomalib 2.6.2 and reject unsupported exports (#606).

Installation

sima-cli neat install apps@v0.6.0

The installer also installs Neat Library v0.6.0 and the latest Neat Insight. If a paired Neat SDK already provides a matching Neat Library and Insight, skip them with NEAT_APPS_SKIP_DEPENDENCIES=1 sima-cli neat install apps@v0.6.0 (#597).

Compatibility

Component	Version
SiMa.ai Platform	3.0.0
Neat Library	v0.6.0 (4404f3fd5a91)
Neat Insight	v0.0.9

Apps 0.6.0 downloads models from Model Zoo 3.0.0. Rotated boxes, MJPEG channels, the BlazePose 3D view and the GenAI tab need Neat Insight v0.0.9.

Migration notes

• Apps v0.6.0 requires Platform 3.0.0 and Neat Library v0.6.0. Rebuild native applications built from source against Neat Library v0.6.0.
• Renamed or replaced examples:
  • image-classifier → image-classification-explorer (#520)
  • multi-stream-people-tracker → multi-stream-tracker (#538)
  • yolov8-instance-segmenter → single-stream-instance-segmenter with model.family: yolov8 (#537)
• High-density profiles are renamed to config-30x720p20fps.yaml and config-60x720p10fps.yaml, and the default profile has 20 streams. The PCIe loader accepts exactly 20, 30 or 60 streams (#618).
• Single-stream object detection and instance segmentation read RTSP settings from source.rtsp. Move flat source.url and source.codec keys there (#644).
• RF-DETR detection downloads rf_detr_<size>_w8a8_w16a16 model pairs, and raw_video_max_fps is removed (#544, #638).
• Model downloads use sima-cli download from Model Zoo 3.0.0 instead of sima-cli modelzoo get (#616). A first download may ask for a Developer Portal login (#663).
• GenAI Studio is backend-only: its web page, audio playground and terminal CLI are removed, and Insight's GenAI tab replaces them. The Studio's source moves from src/python/ui to src/python/api, and the first run.sh start migrates existing voices, documents and logs (#602, #629). The Qwen3-VL model is now simaai/Qwen3-VL-4B-Instruct-Autoround-a16w4 (#621).
• The PCIe example installs the host package from core/pciehost/...@v0.6.0 (#595).

Contributors

@Onat-Inak, @Yukkta-Seelam, @florianvoss-commit, @nicholasxu-wq, @michaelschlosserx, @shanguangzai, @smaurya01, @bandaruprudhviram, @bhavyasharma-sima, @manuel-roldan, @sairam5096

Full Changelog: v0.5.1...v0.6.0