# System Requirements This page lists the hardware, software, and platform requirements, and the supported models, for setting up and running the application. ## Software and Hardware Requirements ::::{tab-set} :sync-group: os :::{tab-item} Windows :sync: windows - **OS**: Windows 11 - **Recommended processor**: Intel® Core™ Ultra Series 1, 2, 3, and 4 processors (with integrated GPU support) - **Memory**: 32 GB RAM (minimum recommended) - **Storage**: At least 50 GB free (for models and logs) - **GPU/Accelerator**: Intel® iGPU (Intel® Core™ Ultra Series 1, Intel® Arc™ GPU, or higher) for summarization acceleration - **NPU**: Intel® NPU (Intel® Core™ Ultra Series 1 or higher) for Video pipelines - **NPU Driver**: Download and install the latest version from [Intel NPU Driver Download Page](https://www.intel.com/content/www/us/en/download/794734/intel-npu-driver-windows.html) - **Python**: 3.12 (the desktop app's **Setup** screen can install it) - **Node.js**: Not needed; the desktop app bundles its own. Building the installer from source needs Node.js 20.19+ or 22.12+ ::: :::{tab-item} Linux :sync: linux - **OS**: Ubuntu 24.04 LTS - **Recommended processor**: Intel® Core™ Ultra Series 1, 2, 3, and 4 processors (with integrated GPU support) - **Memory**: 32 GB RAM (minimum recommended) - **Storage**: At least 50 GB free (for models, Docker images, and logs) - **GPU/Accelerator**: Intel® iGPU (Intel® Core™ Ultra Series 1, Intel® Arc™ GPU, or higher) required for Content Search reranking and used for summarization acceleration - **NPU**: Intel® NPU (Intel® Core™ Ultra Series 1 or higher) for Video pipelines - **GPU/NPU Drivers**: Install the latest Intel GPU and NPU drivers from [Intel NPU Driver releases](https://github.com/intel/linux-npu-driver/releases) and the [Intel GPU compute runtime](https://github.com/intel/compute-runtime/releases). The Linux app container installs its pinned OpenCL/Level Zero userspace runtime and needs host GPU access through `/dev/dri`. - **Docker**: Docker Engine and Docker Compose plugin - **Python**: 3.12 - **Node.js**: v18+ (for frontend) ::: :::: ## Audio Pipeline Supported Models ### ASR (Automatic Speech Recognition) - **Whisper (all models supported)** - Recommended: `whisper-small` or lower for CPU efficiency - Runs on **CPU** (Whisper is CPU-centric) - **FunASR (Paraformer)** - Recommended for **Chinese transcription** (`paraformer-zh`) - Supports transcription of .mp3/.wav audio files up to 45 minutes long. ### Summarization (LLMs) - **Qwen Models (OpenVINO™)** - `Qwen3-VL-8B-Instruct` (default, shared vision-language model) - `Qwen3.5-9B` - Summarization supports up to 5,120 tokens on GPU - Run summarization on **GPU** (Intel® iGPU / Intel® Arc™ GPU) for faster performance. ### Content Segmentation and Topic Search - **Embedding Model**: `intfloat/multilingual-e5-small` for semantic topic indexing and search - **Vector Store**: ChromaDB - **Content segmentation**: uses the shared vision-language model (`Qwen3-VL-8B-Instruct` by default) ### Supported Weight Formats - **int4** → Recommended default (fast + memory-efficient on GPU) - **int8** → Higher accuracy, larger memory footprint - **fp16** → Full precision (use only if sufficient memory is available) ## Video Analytics Pipeline - Supports 3 concurrent video pipelines (front, back, content) up to 45 minutes - Supports .mp4 format and RTSP streams - Outputs processed video via RTSP and HLS/WebRTC streaming (MediaMTX) For pipeline architecture and processing stages, see [How It Works](../how-it-works.md#video-analytics-pipeline). ### Supported Models | Model | Format | Used In | Purpose | | :--- | :--- | :--- | :--- | | **YOLOv8m-pose** (default) | OpenVINO™ IR | Front pipeline | Person detection + 17-keypoint pose estimation | | **YOLOv8s-pose** (default) | OpenVINO™ IR | Back pipeline | Lightweight person detection + pose estimation | | **YOLO11m/s-pose** | OpenVINO™ IR | Front / Back pipeline | Alternative pose model | | **YOLO26m/s-pose** | OpenVINO™ IR | Front / Back pipeline | Alternative pose model | | **ResNet-18** | OpenVINO™ IR | Content pipeline (front and back pipelines on Linux only) | Content frame classification | | **MobileNet-V2** | OpenVINO™ IR | Front pipeline (Linux only) | Lightweight classification | | **Person-ReID-retail-0288** | OpenVINO™ IR | Front pipeline | Person re-identification and tracking | - All models run in OpenVINO™ Intermediate Representation (IR) format - Inference supported on **CPU**, **GPU**, and **NPU** (configurable per pipeline) - Default inference device: **NPU** (recommended for best performance on Intel® Core™ Ultra processors) ## Content Search Pipeline ### Content Search Supported Models | Model | Purpose | Device | | :--- | :--- | :--- | | **Qwen3-VL-8B-Instruct** | Vision Language Model for video summarization and Q&A | GPU | | **CLIP/clip-xlm-roberta-base-vit-b-32** | Visual embedding for images and video frames | CPU | | **intfloat/multilingual-e5-small** | Text embedding for document chunks | CPU | | **BAAI/bge-reranker-base** | Cross-encoder reranking for search results | GPU | ### Supported File Formats | Category | Extensions | | :--- | :--- | | **Video** | `.mp4` | | **Document** | `.txt`, `.pdf`, `.docx`, `.doc`, `.pptx`, `.ppt`, `.xlsx`, `.xls`, `.html`, `.htm`, `.xml`, `.md` | | **Image** | `.jpg`, `.jpeg`, `.png` |