BMRT and RKNN backends
Chip-aware model and package validation prevents artifacts from being treated as interchangeable.
Deploy video
AI as field-ready
edge systems.
Connect video streams, models, rules, and system output in one C++ edge runtime built for real deployments.
v1.1.0 stable4 target-device archivesApache 2.0C++17 runtimeSophon · RKNN · x86
v1.1.0 stable release
One workflow across Sophon, Rockchip RKNN, and x86.
v1.1.0 ships the common runtime and operations model while keeping every package, model artifact, media profile, and validation result bound to its target platform.
BM1688CV186XRK3576RV1126Bx86macOS Preview
Persistent staged uploads
Resume, cancel, admit by live resources, and clean temporary data deterministically.
Open and Protected paths
Open packages retain plaintext user models; Model Guard 2.3 remains a controlled commercial path.
Multi-platform ScenarioBench
Controlled results expose exact workloads and limits without turning short runs into production sizing.
Field AI infrastructure
Not just algorithms. A reusable field AI base.
Real deployments need video ingestion, decoding, model execution, rules, live preview, alarms, snapshots, and system integration.
RTSPIPCNVR
CosmoEdgeEdge runtime
Model inference
Rule orchestration
System output
OSD overlayrule eventevidence frame{ api: true }See the complete engine workflow.
From task creation through live preview to downstream delivery, the videos show how the engine runs as one console.
Visual pipeline
Connect video ingestion, inference, rules, and reporting as one task.
Prompt-driven AI
Bring GroundingDINO and VLM-based visual judgment to edge-side workflows.
Configurable applications, not one-off projects.
Video access, model inference, scene tasks, alarms, display, and output are modularized so teams can deliver, adjust, and replicate projects faster.
INBuild without rebuilding the stackConfigure inputs, models, rules, overlays, and delivery inside one workflow.
TPLSave pilots as reusable templatesOnce a scenario works, its rules, parameters, and outputs can be reused across more sites.
ITRTurn field feedback into iterationUse alarm review and task configuration to tune recognition, notifications, and operating rules over time.
Describe the task
before training another model.
Built-in multimodal capabilities help teams validate long-tail field scenarios with prompts before committing to new data collection and training cycles.
01Describe the targetDINO open-vocabulary detection quickly validates long-tail objects such as foreign items, debris, or temporary construction materials.
02Judge the visual stateVLM prompts inspect equipment states, human behavior, site conditions, and semantic visual changes.
03Turn perception into an eventRules, zones, snapshots, and MQTT / Webhook delivery make prompt-driven results reviewable and reportable.
Designed to be deployed, monitored, and integrated.
CosmoEdge is built for long-running edge systems, not one-off demos. It gives teams a place to manage running tasks, track resource usage, review alarm history, and connect downstream systems.
Multi-camera tasksRuntime overviewModel resourcesAlarm recordsMQTT / WebhookEdge hardware
Quick Start
Start on x86, deploy to edge hardware when ready.
Use x86 developer mode as the shortest validation path, then follow the full tutorial for commands, screenshots, and checks.
Developer path
Validate the runtime, then move into the full tutorial.
Bring up the local stack, open the console, create one scenario task, and confirm event delivery before moving the same workflow toward Sophon BM1688 deployment.
Decoupled layered architecture.
Algorithms, applications, and hardware can iterate independently, then converge in one edge intelligence hub.
CosmoEdge Edge Intelligence Hub
DeliverableRepeatableContinuously Evolving
CosmoEdge Ready Device
Hardware, models, and delivery support in one package.
Open source gives you the engine, web console, and workflow.
Certified device packages add preconfigured Sophon BM1688 acceleration, production models, scenario presets, and field deployment support for common tasks such as person detection, vehicle analytics, helmet recognition, fire / smoke, intrusion, line crossing, and integration pilots.
Sophon BM1688 NPU
Multi-channel video
LAN / USB / HDMI / I/O
Local models and records
Field-ready power
16-channel CV benchmarkThe open-source engine is available today and does not require a device purchase. For certified configurations, availability outside mainland China, or project deployment support, contact hello@cosmowander.ai.
Built for solution teams turning vision AI into field systems.
Safety monitoring
Detect helmets, intrusion, fire, smoke, and zone violations with traceable records.
RTSP -> YOLO -> Zone Rule -> MQTT AlertTraffic and flow analytics
Track people, vehicles, line crossing, occupancy, and multi-camera flows.
Camera -> Tracker -> Line Rule -> Count OutputVisual inspection
Use VLM prompts to inspect equipment states, site conditions, and long-tail scenarios.
Image -> VLM Prompt -> State Rule -> EventSystem integration
Deliver snapshots, alarms, and structured payloads into downstream business systems.
Detector -> JSON Payload -> Webhookv1.1.0 is available now
Choose your target, verify the archive, and deploy.
CosmoEdge v1.1.0 is the latest stable public release. The release center binds each device archive to its target chip and SHA-256 while keeping controlled benchmark evidence within its own qualification boundary.