Turn video AI models into deployable edge applications — a C++ edge AI engine for Sophon, Rockchip, and x86.
Build and operate video analytics, VLM, and event workflows through a consistent orchestration experience. Each platform uses its own runtime, build, and model artifacts.
Start Here · Recent Development · Quick Start · Platforms · Validation · Documentation · 简体中文
hero1_compressed.mp4
CosmoEdge goes beyond model serving with a complete application layer for model import, visual orchestration, alarms, and event delivery. The core engine and console in this repository are released under Apache-2.0; certified hardware, commercial preset models, and Model Guard distribution protection have separate boundaries.
| Your task | Start here | Project and version |
|---|---|---|
| Deploy video AI applications, configure models, scenes, and alarms | Quick Start · Platforms | Engine and console in this repository; stable v1.1.0 |
| Customize an object detector from your data | CosmoEdge Training Skill | Independent project; published v0.1.0-preview.1, with later revisions on its main |
| Let an AI assistant query alarms, retrieve images, and request confirmed detection start/stop | CosmoEdge Connect | Independent local MCP / WorkBuddy integration; v0.1.0-alpha.1 source preview |
The training Skill and Connect are installed and released separately. See Companion Projects for prerequisites, validation scope, and version links.
Stable release: v1.1.0. Covers BM1688, CV186X, RK3576, RV1126B, and x86, with an Apple Silicon macOS Docker Preview. See the v1.1 report for multi-platform measurements and controlled 72-hour fixed-profile results, and the GitHub Release for installation packages.
Current main: development since v1.1.0. Merged changes include:
- Video access: ONVIF discovery and authenticated access, plus GB28181 device management with TCP/UDP signaling and video access.
- Models and visual workflows: YOLO26 and oriented bounding-box (OBB) handling; native BM1688 Laya-V model integration, alarm review, and multi-ROI question configuration for images and video. Implementation and completed validation are described in PR #219.
- Platform integration: Management APIs for versioned resource delivery, resumable uploads, task start/stop, and state readback.
- Deployment and engineering quality: RK3576 protected-model handling, model-package validation, video-sample event measurement and VLM evaluation tools, and fixes to task recovery, media handling, and log retention. See the Unreleased changelog.
These main-branch changes are not included in v1.1.0 installation packages. Consult each feature's documentation for required models, target platforms, and validation scope. The v1.1 benchmarks remain evidence for their recorded candidates and conditions, not performance measurements of later main builds or new models.
CosmoEdge provides one engine architecture and orchestration experience, but each build selects one inference backend and uses models generated for that target platform. The table describes the v1.1.0 release scope and its associated evidence.
| Platform | Status | Runtime / model artifact | Current scope and evidence |
|---|---|---|---|
| Sophon BM1688 | v1.1 supported / primary; VLM supported | BMRT / .nn |
Validated VLM at 0.1 FPS/ch through 6 channels, plus v1.1 workload evidence |
| Rockchip RK3576 / RV1126B | v1.1 supported; RK3576 VLM supported | RKNN / target-specific .rknn; RKLLM on RK3576 |
RK3576 validates VLM at 0.1 FPS/ch through 4 channels; RV1126B VLM is outside this release claim. Other platform-specific results remain separate in the v1.1 report |
| Sophon CV186X | v1.1 supported; VLM supported | BMRT / target-specific .nn |
Validated VLM at 0.1 FPS/ch through 6 channels, plus model-import and workload evidence in the v1.1 benchmark |
| x86 Linux / Windows; Apple Silicon macOS | Linux / Windows supported; macOS Preview | ONNX Runtime / .onnx |
Mac uses amd64 emulation for one local-video developer workflow, not native performance evidence |
| Sophon BM1684X | Planned | — | Not part of the current release scope |
The commands below clone and build current main. For the stable release, use the v1.1.0 packages, or check out the v1.1.0 tag before building and follow that version's documentation.
No edge hardware is required. The x86 mode uses the same UI and workflow with lower throughput than an NPU deployment.
# 1. Clone
git clone https://github.com/cosmo-wander-ai/cosmo-edge.git
cd cosmo-edge
# 2. Start on Linux
sudo docker compose -f docker-compose.x86.yml up -d --build
# Windows: docker compose -f docker-compose.x86.windows.yml up -d --build
# 3. Open http://localhost:8080Apple Silicon macOS uses a separate amd64-emulation Preview path:
./scripts/macos-docker-preview.sh doctor
./scripts/macos-docker-preview.sh up
./scripts/macos-docker-preview.sh status
# Open http://127.0.0.1:8080The Mac Preview provides a local, single-video workflow under linux/amd64
emulation on Apple Silicon. Read the full
macOS Docker Preview scope first.
After startup, use the Scenario Configuration tutorial to create your first AI detection task. Docker Compose V1 users can replace docker compose with docker-compose.
git clone https://github.com/cosmo-wander-ai/cosmo-edge.git
cd cosmo-edge
# BM1688 (default when the chip model is omitted)
./scripts/docker-compose.sh -f docker-compose.sophon.yml run --rm cosmo-sophon-package --chip bm1688
# CV186X
./scripts/docker-compose.sh -f docker-compose.sophon.yml run --rm cosmo-sophon-package --chip cv186xThe wrapper selects the available Compose implementation and requests elevated
Docker access only when required. The build selects the matching model resources
without a model-path argument and exports each target independently under
build_output/public-runtime/<chip>/, together with TARGET_CHIP and
SHA256SUMS.
The default Open package contains plaintext models and requires no device authorization. Verify the chip marker and checksum before deployment. The Build Guide is the authoritative build reference; follow the Deployment Guide for SSH installation, web upgrade, recovery, and post-reboot acceptance.
./scripts/docker-compose.sh -f docker-compose.rockchip.yml pull cosmo-rockchip-package
# RK3576 (uses the tracked RK3576 model resources)
COSMO_TARGET_CHIP=rk3576 ./scripts/docker-compose.sh \
-f docker-compose.rockchip.yml run --rm cosmo-rockchip-package
ls -lh build_output/rk3576/
# RV1126B (requires a prepared target model overlay)
COSMO_TARGET_CHIP=rv1126b ./scripts/docker-compose.sh \
-f docker-compose.rockchip.yml run --rm cosmo-rockchip-package
ls -lh build_output/rv1126b/One digest-pinned Rockchip builder shares the compiler and RKNN SDK while
selecting an isolated MPP/RGA profile for each chip. RKLLM v1.3.0 is required
and packaged only for RK3576. Target markers, media-profile identities, and
checksums are exported under build_output/<chip>/. See the
Build Guide and
RK3576 integration guide for model
and device-validation boundaries.
For CV186X, follow the CV186X Quick Start for package installation, model import, first-event verification, upgrade, and recovery boundaries.
- Real-time video analytics: detection, classification, tracking, zones, counters, OSD, and alarm snapshots.
- Prompt-driven vision: VLM state judgment and GroundingDINO open-vocabulary detection alongside conventional CV pipelines.
- Visual application workflows: connect models, rules, events, and output actions in the browser.
- Edge integrations: operate managed tasks and deliver structured events through REST, WebSocket, MQTT, or HTTP webhooks.
| Capability | What it covers | Go deeper |
|---|---|---|
| Native runtime | C++17 engine for multi-channel media, inference scheduling, OSD, tasks, and events | Architecture |
| Visual orchestration | Browser-based pipeline composition, task binding, parameter validation, and live feedback | Pipeline tutorial |
| Inference and media | Platform backends for Sophon, RKNN, and x86; platform-specific builds and model artifacts | Build Guide |
| VLM and DINO | Prompt-based judgment, open-vocabulary detection, and optional VLM review before a detection alarm is reported | VLM Guide |
| Operations and integration | Model management, alarms, event history, REST, WebSocket, MQTT, and webhooks | API Overview |
| Model onboarding and protection | Model conversion, import, validation, and the Open/Protected distribution boundary | Model Porting Guide |
▶ Watch: compose a complete visual pipeline
pipline2_compressed.mp4
▶ Watch: GroundingDINO and VLM visual workflows
promot1_compressed.mp4
Already have a model-porting, integration, or UI task? Give your usual coding agent the business goal, available materials, target device or test environment, and acceptance criteria. The repository provides task entry points, examples, checks, and evidence boundaries so the result can include importable artifacts, scoped code changes, and a verifiable conclusion.
Start with Agent-Assisted Development, then use the Model Porting Guide or Contributor Guide for the task at hand.
The v1.1 report covers BM1688, CV186X, RK3576, and RV1126B. It includes 49 independent small-model case reports for person detection, no-safety-helmet analysis, and a concurrent mixed workload at 5 FPS per business task, a controlled 72-hour dual-CV configured profile, and validated VLM performance for BM1688, CV186X, and RK3576.
- English benchmark index
- 中文基准报告索引
- English primary report (rendered documentation site)
- 72-hour dual-CV report (rendered documentation site)
- Methodology and reproduction
Expand v1.1 candidate identities, performance tables, and test conditions
The short-run capacity refresh uses source commit 89c73a7464a81ef378686447d7c1eeb88b988455, tree 6857fbcce72c7af64e6cb23a27e66a405e9df9af, one fixed 1080p24 input, 30-second steps, and the gates recorded in the release manifest. The 72-hour observation uses source commit 44209759f450e96cda265acfa8bc6d17a1138888, tree 5cbdefeaefaf642407356c22c271ccc7d57935b0, and the same controlled input.
| Platform | 72-hour configured channels | Task bindings | Sample coverage | Min / avg FPS | Result |
|---|---|---|---|---|---|
| BM1688 | 8 | 16 | 4316 / 4320 | 4.68 / 5.086 | PASS |
| CV186X | 8 | 16 | 4316 / 4320 | 4.54 / 5.085 | PASS |
| RK3576 | 8 | 16 | 4316 / 4320 | 5.00 / 5.098 | PASS |
| RV1126B | 4 | 8 | 4316 / 4320 | 4.85 / 5.230 | PASS |
These results describe the tested fixed-channel profiles under controlled local-loop input; they are not maximum-capacity or RTSP-resilience measurements. Execution details are documented in the methodology.
Each channel concurrently runs two business tasks across three model stages: person detection has one detector stage, while no-safety-helmet analysis has a detector followed by a classifier.
| Platform | Workload per channel | Model stages/ch | Target FPS/task | Passing channels | Business-task bindings |
|---|---|---|---|---|---|
| BM1688 | Person detection + no-safety-helmet analysis | 3 | 5 | ≥16 | 32/32 |
| CV186X | Person detection + no-safety-helmet analysis | 3 | 5 | ≥8 | 16/16 |
| RK3576 | Person detection + no-safety-helmet analysis | 3 | 5 | ≥8 | 16/16 |
| RV1126B | Person detection + no-safety-helmet analysis | 3 | 5 | ≥4 | 8/8 |
The three VLM platforms use one final protocol: each newly added route completes task-local readiness before a 60-second measurement window, and every active route must pass the executed 80% threshold against the 0.1 FPS-per-channel target. Each platform's fixed candidate package also passed model load, task creation, valid inference, event/alarm output, and task recovery after service restart; these claims do not extend to later rebuilt packages.
| Platform | Target FPS/ch | Last passing channels | First failure |
|---|---|---|---|
| BM1688 | 0.1 | 6 channels | 7-channel FPS ratio 69.5%, below the 80% gate |
| CV186X | 0.1 | 6 channels | 7-channel FPS ratio 69.2%, below the 80% gate |
| RK3576 | 0.1 | 4 channels | 5-channel FPS ratio 69.7%, below the 80% gate |
| RV1126B | — | — | Outside this VLM validation |
The capacity values are exact short-run gate boundaries under the recorded conditions—not maximum-capacity certification or production recommendations. The separate 72-hour table verifies only its configured profile. See the v1.1 benchmark for the complete matrices. Previously published data remains available through one v1.0 historical archive link.
+------------------------------------------------------------------+
| Web Console | Visual Orchestration | REST / WebSocket / MQTT |
+--------------------------------+---------------------------------+
|
+--------------------------------v---------------------------------+
| C++ Engine Core |
| Media | Inference | Tasks | Rules | Alarms | Events | Models |
+--------------------------------+---------------------------------+
|
+--------------------------------v---------------------------------+
| Inference and Media Backend Interfaces |
+--------------------+----------------------+----------------------+
| Sophon BMRT/VPU | RKNN + MPP/RGA | ONNX Runtime/FFmpeg |
| BM1688; CV186X | RK3576 | x86 Linux / Windows |
+--------------------+----------------------+----------------------+
One build selects one inference backend. Model artifacts are generated for the target platform, and feature/model coverage and capacity remain platform-specific. The v1.1.0 Model Guard Protected distribution notes apply to Sophon packages; current main also includes RK3576 protected-model handling, described in PR #151.
| Start here | Best for |
|---|---|
| Documentation Home | Full documentation index and learning path |
| Quick Start Guide | First setup and scenario run |
| Scenario Configuration | Building scene-level workflows |
| VLM Guide | Prompt-based visual judgment and events |
| Model Porting Guide | Importing your own model |
| Agent-Assisted Development | Delegating an extension task with verifiable results |
| Companion Projects | Training Skill and Connect / MCP prerequisites and independent versions |
| Changelog | Release history and unreleased main-branch changes |
| Build Guide | x86, Sophon, and RK3576 build/package paths |
| API Overview | REST, WebSocket, MQTT, and webhook integration |
Certified devices add preconfigured acceleration, validated commercial model packages, and dedicated deployment support; they do not unlock separate software features. Devices are available in mainland China from the Taobao store; contact [email protected] for other regions or project support.
Contributions are welcome through scoped bug reports, documentation improvements, scenarios, and integration notes. Read CONTRIBUTING.md before opening a pull request.
GitHub Discussions is the official searchable English Q&A and community-support channel for v1.1, and general reproducible defects belong in GitHub Issues. Users in mainland China can use Gitee Issues for repository access, installation, usage, release downloads, and device-adaptation questions. Code changes and pull requests remain on GitHub; report vulnerabilities privately through SECURITY.md. No Discord channel is operated for this release.
Can I try CosmoEdge without Sophon or Rockchip hardware?
Yes. Use x86 developer mode on Linux or Windows. Apple Silicon Macs can use the Docker Preview for the console, pipeline, model-management, and integration workflow with one local video. The Mac path uses amd64 emulation and does not enable Model Guard. Edge NPU hardware is still required for target-platform acceleration and capacity validation.
What is the boundary between the Open and Protected packages?
They expose the same application features and use the same MD5 upgrade lifecycle. Open uses plaintext models without device authorization; Sophon Protected packages can carry encrypted commercial preset models and provisioning tooling that require a device-bound certificate. Application archives themselves are not signed.
Can I use my own trained models?
Yes. For training from data, use the independent Training Skill. For an existing model, use the model-porting path to validate the tensor, preprocessing, post-processing, target runtime, and business accuracy contract. A model artifact must be generated for the platform where it will run.
How production-ready is CosmoEdge?
v1.1.0 is the multi-platform release line for BM1688, CV186X, RK3576, RV1126B, and x86, with VLM support validated on BM1688, CV186X, and RK3576 plus a scoped macOS Docker Preview. The linked report records measured workload boundaries, validated VLM performance, and controlled 72-hour fixed-profile results. Production sizing still requires validation with your own models, streams, accuracy requirements, and deployment conditions.
CosmoEdge is licensed under the Apache License 2.0. Copyright 2026 CosmoEdge Contributors.
An open-source project by Cosmo Wander AI and the CosmoEdge contributors.
Turn video AI models into deployable edge applications.
📦 GitHub remains the development and pull-request source of truth. Gitee automatically mirrors the code and provides release access and a Chinese issue channel for users in mainland China. See MIRRORING.md.