CosmoEdge

English

简体中文

English

简体中文

Appearance

中文

CosmoEdge / Ready devices

Models included.
Ready for video.

Preinstalled models cover safety helmets, fire and smoke, falls, and people and vehicles. Connect video and configure a scenario with models adapted for your platform.

21–27model packages, by platform
29configurable scenario templates
Explore included models Compare devices
See multi-stream and 72-hour test results Have your own hardware? Explore self-deployment
Sophon BM1688
CosmoEdge BM1688 device, front view
BM1688 model shown. Ports follow the selected device configuration.8 GB · 64 GB eMMC
Safety helmetsFire & smokeFallsPeople & vehicles

Models, the runtime and scenario presets ship with the device. Adjust regions, thresholds and event rules for your site.

For international purchases, contact us by email.

Configuration guide

Compare platforms, memory and NPU compute.

All three configurations include 64 GB eMMC storage. Start with your model format, memory needs and workload.

01

Rockchip

RK3576

Rockchip RKNN

People, vehicle and safety analysis, using the available RKNN models and scenario presets.

Nominal NPU compute
6 TOPS
Memory
4 GB LPDDR5
Storage
64 GB eMMC
Model packages
21 included
Ask about purchasing Models and scenarios
02

Sophon

CV186AH

Sophon · 8 GB memory

Uses CV186X platform resources, including text-prompted detection and image segmentation.

Nominal NPU compute
7.2 TOPS
Memory
8 GB LPDDR4X
Storage
64 GB eMMC
Model packages
26 included
Ask about purchasing Models and scenarios
03

Sophon

BM1688

Sophon · 16 TOPS

27 model packages. Use the published mixed-workload results to assess a multi-task deployment.

Nominal NPU compute
16 TOPS
Memory
8 GB LPDDR4X
Storage
64 GB eMMC
Model packages
27 included
Ask about purchasing Models and scenarios

TOPS is nominal hardware compute. Stream capacity and speed depend on the model, input frequency and runtime configuration. CV186AH uses the CV186X resource platform below.

Measured workloads

Measured performance for person and safety-helmet analysis.

Person detection + no-safety-helmet analysis: two tasks and three model stages per stream, targeting 5 FPS per task. These are the platform results in the published report.

Test platformStreams passing the mixed-workload test72-hour fixed workloadModel resource differences
RK3576≥ 88 streams · 16 tasksPeople, vehicles, safety, vision-language
CV186X≥ 88 streams · 16 tasksAlso includes text-prompted detection and segmentation
BM1688≥ 168 streams · 16 tasksAlso includes text-prompted detection and segmentation

Mixed-workload tests use fixed 1080p / 24 FPS local looped video, 30 seconds per level and no preview client. ≥ means the test ceiling passed. The 72-hour run is a separate fixed-stream observation. See the report for platform, model and test details.

Read the full report

What’s included

Your device, software and included resources.

01

Device and runtime

Your device comes with CosmoEdge and the NPU runtime configured for its chip.

02

Models and scenario presets

Platform-specific model packages and 29 configurable scenario templates. Connect an input and adjust regions, thresholds and event rules for your site.

03

Deployment and project scope

Deployment support is available. Agree the scope of custom model adaptation, business integration and on-site services during project discussions.

Included models, templates and services follow the selected listing and delivery agreement. Discuss model integration, site rules and business system needs before purchasing.

Included resources / 2026-09-08

Models and scenario templates by platform.

Browse models and templates by task. To run a template, select compatible models and connect an image or video source.

01 / MODEL CATALOG

BM1688
included models 27

People and vehicle analysis, safety behavior recognition, text-prompted detection, segmentation and vision-language analysis.

Counts refer to model packages. Classification models usually follow a detector; embeddings and landmarks support face or person comparison workflows.

All three platforms include a vision-language model, enabled according to the device delivery configuration.

Person detectionPedestrianDetection
Locate people for region rules, counting and behavior classification.
People and vehicle detectionVehicle
Detect people, non-motor vehicles and motor vehicles.
Helmet classification · head & shouldersHsHelemet
Classify safety helmets, helmets, ordinary hats or no hat from head-and-shoulder crops.
Fire and smoke detectionFireDetection
Detect flames and smoke.
Phone use and smokingPhoneSmoking
Recognize phone calls, phone use and smoking.
Vision-language analysisCosmoEdge-VL-Judge-0.8B
Analyze an image from a prompt and generate a text assessment.
Show 21 more modelsCollapse full model list
Face detectionFaceRGBDetection
Locate faces for comparison, mask recognition and image quality analysis.
Face embeddingsFaceRGBRecognition
Extract features for face database search and comparison.
Mask recognitionFaceMask
Classify faces as wearing or not wearing a mask.
Face quality assessmentFaceQuality
Assess face orientation, completeness and sharpness.
Face landmarksFaceLandmark
Locate facial keypoints for alignment and feature extraction.
Head-and-shoulder detectionHsDetection
Locate head-and-shoulder regions.
Person attributesPedestrianAttribute
Identify clothing, colors and other person attributes.
Person embeddingsPedestrianRecognition
Extract features for person or workwear comparison.
Person image qualityPedestrianQuality
Select person images suitable for comparison.
Sleeping classificationPedSleep
Identify slumped or lying sleeping postures.
Vehicle attributesVehicleAttribute
Identify vehicle type, color and orientation.
E-bike classificationElectricElevatorClassify
Distinguish bicycles from electric bicycles for task orchestration.
Waste pile detectionRubbishDumpDetection
Locate accumulated waste.
Helmet classification · personPedHelemet
Classify safety helmets, helmets, ordinary hats or no hat from person crops.
Fall classificationPedFalldown
Classify detected person targets for falls.
YOLOv8 object detectionYolov8Detect
General object detection for images and video.
YOLOv12 object detectionYolov12
General object detection for images and video.
Rodent detectionMouseDetection
Locate rodents.
YOLO26 object detectionYolo26
Object detection for image analysis.
Text-prompted detectionGrounddingDIno
Locate objects from text prompts with Grounding DINO.
Image segmentationSAM2
Generate object-region masks with SAM2.

02 / SCENARIO TEMPLATES

Shared scenario templates 29

Choose an image or video source, connect the models, then set regions, lines, timing and event conditions.

  • 01
    Region intrusionRaise an event for intrusion into a region.
  • 02
    Region person countCount people in a region in real time.
  • 03
    No safety helmetRaise an event for a person without a safety helmet.
  • 04
    Detect, then classifyLocate an object, then classify the cropped target.
Explore all 29 templatesCollapse scenario templates
Face comparison
1:N face comparison.
Region intrusion
Raise an event for intrusion into a region.
Crowding
Raise an event when the person count exceeds a limit.
Person fall
Raise an event for a detected fall.
Vehicle overstaying
Raise an event when a vehicle exceeds its allowed dwell time.
Smoke detection
Raise an event for persistent smoke.
Flame detection
Raise an event for persistent flames.
Smoking detection
Raise an event for smoking.
Phone call detection
Raise an event for a phone call.
People flow counting
Count entries and exits across a line.
No safety helmet
Raise an event for a person without a safety helmet.
Unattended post
Raise an event when too few people remain in a region for a set duration.
Phone use detection
Raise an event for phone use.
License plate recognition
Locate and read plates; requires a suitable plate recognition model.
Reflective vest comparison
Compare reflective vest appearance.
Region person count
Count people in a region in real time.
Line crossing
Raise an event for one-way or two-way line crossings.
Workwear comparison
Raise events based on a workwear reference library.
Sleeping on duty
Raise events for slumped or lying sleeping postures.
Text-prompted image detection
Locate image objects using text prompts.
Image face recognition
Detect faces, extract features and identify faces in an image.
Text-prompted video detection
Detect prompted objects and analyze events in video.
Image segmentation
Generate segmentation results for image objects.
Video segmentation
Segment objects in video.
Video vision-language analysis
Analyze video frames from prompts and produce assessments.
Image object detection
Detect objects in images.
Detect, then classify
Locate an object, then classify the cropped target.
Detect, then segment
Detect an object, then generate its segmentation result.
Image vision-language analysis
Generate an assessment from an image and a prompt.
Run your first detection →Configure a scenario →

Before you start

Questions before you buy.

Can I use my own YOLO model?

Yes, through the model integration workflow. Match the target format, inputs, outputs and parser configuration. Training weights need export and platform adaptation first.

Read the integration guide →
How do I try a ready-made scenario?

Start with the included models and presets. Run the workflow with sample footage, then use site footage to tune regions, thresholds and rules and confirm the results.

Run your first detection →
Can I use the software without buying a device?

Yes. The engine code is Apache-2.0 licensed and supports self-deployment. Model assets have separate sources and licenses.

Check model sources and licenses →
How do I choose for multiple streams or snapshots?

Selection depends on the models, tasks per stream, input frequency and latency requirements. Share the details you have to discuss a configuration and validation approach.

Discuss device selection →

Your next step

We’ll help you choose a device for your tasks.

Discuss your project →Run your first detection

Engine code: Apache-2.0. Models and dependencies retain their own licenses.

Copyright © CosmoWander AI

CosmoEdge / Contact

Let’s talk about your edge AI project.

Model deployment, hardware selection or system integration—start with the question you’re working on.

Business & project enquiries

hello@cosmowander.ai
Email us

Installation, usage and open-source questionsVisit community discussions
What should I include?

Share what you know; we can work through the remaining details together.

  • Use case

    What to detect and how events will be used

  • Existing setup

    Existing models, hardware or systems

  • Workload

    Stream count, frequency and latency needs

  • Project stage

    Your project stage and immediate questions