# CosmoEdge: C++ Edge AI Engine for Video Analytics on Sophon, Rockchip, and x86 > CosmoEdge is a production-grade C++ edge AI engine that runs video analytics and on-device VLM workloads across Sophon BM1688/CV186X, Rockchip RK3576/RV1126B, and x86 hardware, with a visual orchestration console, real-time OSD, event delivery, and reproducible benchmark artifacts. The core engine is Apache-2.0; commercial preset models and the Model Guard distribution protection layer have separate license terms. **cosmo-wander-ai/cosmo-edge** — Production-grade C++ edge AI engine for video analytics and on-device VLM across Sophon, Rockchip RKNN, and x86, with visual orchestration, real-time OSD, events, and reproducible benchmarks. - Repository: https://github.com/cosmo-wander-ai/cosmo-edge - Website: https://www.cosmowander.ai/ - Stars: 1,241 · Forks: 231 - Language: C - License: Apache-2.0 - Published: 2026-09-16 · Updated: 2026-09-16 · Language: en - Canonical page: https://hysenlabs.com/projects/cosmo-wander-ai-cosmo-edge ## What CosmoEdge Is and Who It Is For Most edge AI inference engines stop at model serving: you load a model, pass it frames, and receive predictions. CosmoEdge adds an application layer on top: a visual orchestration console for configuring detection scenarios, an on-screen display system for overlaying results on video, an event and alarm system for downstream notification, and a benchmarking framework with reproducible artifact checksums. The README describes the project as going beyond model serving with a complete application layer for model import, visual orchestration, alarms, and event delivery. The primary audience is teams deploying video analytics on edge hardware from Sophon (BM1688 and CV186X chips) and Rockchip (RK3576 and RV1126B chips). These are NPU-based platforms common in industrial cameras, smart city infrastructure, and robotics applications. CosmoEdge provides platform-specific Docker builds, model artifact pipelines, and inference paths for each target, while keeping the orchestration console and event system consistent across platforms. An x86 Linux and Windows path is included for development and testing without edge hardware. An Apple Silicon macOS path exists as a Preview using amd64 emulation, documented as a single-video developer workflow rather than a performance deployment. ## Platform Support and Inference Backends CosmoEdge uses different inference backends per platform. Sophon chips use BMRT with .nn model artifacts. Rockchip chips use RKNN with target-specific .rknn artifacts, plus RKLLM on RK3576 for VLM workloads. The x86 path uses ONNX Runtime with .onnx models. The README documents the v1.1 platform support in a table. BM1688 and CV186X are listed as supported with VLM validated. RK3576 is supported with VLM validated at 0.1 FPS/ch through 4 channels. RV1126B is supported but VLM is outside the release claim. x86 Linux and Windows are supported. macOS on Apple Silicon is a Preview. Sophon BM1684X is listed as Planned. Version 1.1.0 was released on 2026-08-24. Version 1.0.0 was released on 2026-07-03. The last push to the repository was on 2026-09-24. The project also supports GB28181, the Chinese national standard for video surveillance network systems, with a separate Docker Compose file (docker-compose.gb28181.yml). ## Getting Started on x86 and macOS For x86 Linux, the README's quick start requires Docker and Docker Compose: ```bash git clone https://github.com/cosmo-wander-ai/cosmo-edge.git cd cosmo-edge sudo docker compose -f docker-compose.x86.yml up -d --build ``` For Windows, replace the file flag with docker-compose.x86.windows.yml. After startup, open http://localhost:8080 to reach the orchestration console. On Apple Silicon macOS, the README provides a separate preview script: ```bash ./scripts/macos-docker-preview.sh doctor ./scripts/macos-docker-preview.sh up ./scripts/macos-docker-preview.sh status ``` The Mac preview runs under linux/amd64 emulation and opens at http://127.0.0.1:8080. The README is explicit that this path is isolated to one local-video workflow and does not provide native Apple Silicon performance. It describes it as validated through multiple end-to-end lab rounds, but scoped to a single-video developer use case. After startup, the README directs new users to the Scenario Configuration tutorial on the project documentation site to create the first AI detection task. Docker Compose V1 users can replace docker compose with docker-compose. The repository includes a docker-compose.gb28181.yml file for GB28181 protocol support, the Chinese national standard for network video surveillance. Teams building systems that need to integrate with Chinese municipal or enterprise surveillance infrastructure can use this path alongside the standard x86 or edge builds. The repository is also mirrored on Gitee at gitee.com/cosmo-wander-ai/cosmo-edge for users in regions where GitHub access is intermittent, with the MIRRORING.md file at the repository root documenting the mirror policy and synchronization schedule. An AGENTS.md at the repository root provides guidance for AI coding agents working on the codebase, covering the build system and how to run validation scripts. ## Building for Sophon and Rockchip Hardware Production deployment on Sophon hardware uses a Docker-based cross-build that selects the chip target: ```bash ./scripts/docker-compose.sh -f docker-compose.sophon.yml run --rm cosmo-sophon-package --chip bm1688 ``` For CV186X, replace --chip bm1688 with --chip cv186x. The build exports chip-specific packages to build_output/public-runtime/chip/, including TARGET_CHIP and SHA256SUMS files for verification. The Rockchip build uses a separate Compose file: ```bash ./scripts/docker-compose.sh -f docker-compose.rockchip.yml pull cosmo-rockchip-package ``` The README recommends reading the Build Guide and Deployment Guide in the project documentation before deploying to actual hardware. The deployment guide covers SSH installation, web upgrade, recovery, and post-reboot acceptance procedures. The wrapper script (docker-compose.sh) selects the available Compose implementation (V1 or V2) and requests elevated Docker access only when required. The default Open package contains plaintext models and requires no device authorization; the Protected package uses Model Guard encryption and requires device provisioning. ## VLM Support, Benchmark Evidence, and Model Guard VLM (vision-language model) inference is one of CosmoEdge's more recent capabilities. For Sophon BM1688 and CV186X, VLM is validated at 0.1 FPS/ch through 6 channels each. For Rockchip RK3576, VLM is validated at 0.1 FPS/ch through 4 channels. The v1.1 benchmark pack covers single workloads, concurrent mixed workloads, a controlled 72-hour dual-CV profile, and validated VLM performance. The README emphasizes that benchmark artifacts are sanitized and traceable: each record includes an exact Open-package and running-engine binding, and the SHA256SUMS allow independent verification. VLM figures in the benchmark are described as exact short-run gate boundaries for the recorded protocol, not extrapolated throughput claims. Model Guard 2.3 is the commercial model distribution protection component. It protects preset-model distribution in Sophon Protected packages using model encryption and device-provisioning tooling. The README explicitly states that Open and Protected packages expose the same application features with no SKU-gated software functionality; the difference is model encryption and provisioning tools, not access to orchestration or event features. ## Limitations and Comparison to Nvidia DeepStream The 0.1 FPS per channel VLM throughput is a concrete ceiling for the validated configurations. Applications requiring real-time VLM inference at multiple frames per second on the listed hardware are outside the documented operating envelope. The macOS Preview path runs under amd64 emulation, which means performance on Apple Silicon reflects the emulation layer rather than native capability. The README states this path was validated for one local-video workflow, not for broader workload characterization. BM1684X support is listed as Planned with no schedule. Teams targeting that chip cannot use CosmoEdge v1.1 and must wait for a future release or contribute support. RV1126B VLM is explicitly outside the v1.1 release claim. The README's platform table marks VLM as supported on BM1688, CV186X, and RK3576 only. Teams choosing RV1126B for a project that requires on-device VLM should note this boundary before committing to the platform. The closest comparable project for NVIDIA hardware is NVIDIA DeepStream, which provides a similar application-layer framework for video analytics on NVIDIA GPUs and Jetson devices. DeepStream uses TensorRT as its inference backend and targets CUDA-capable hardware exclusively. CosmoEdge's distinction is its support for Sophon and Rockchip NPU platforms, which are commonly used in cost-constrained industrial and embedded deployments where NVIDIA hardware is not available or not cost-effective. The core engine and console are Apache-2.0. Commercial preset models and the Model Guard component have separate license terms, which the NOTICE and LICENSE files document. The repository includes CODING_STYLE.md, CONTRIBUTING.md, a SECURITY.md, and .clang-format and .clang-tidy configuration, indicating the C++ codebase has active quality standards. A MIRRORING.md file and a Gitee mirror (gitee.com/cosmo-wander-ai/cosmo-edge) are noted in the README for users in regions where GitHub access is limited. ## Conclusion CosmoEdge is the right choice for teams building video analytics applications on Sophon or Rockchip NPU hardware who need a complete application layer rather than just a model serving stack. The Apache-2.0 core engine is freely usable and the benchmark evidence is detailed and reproducible. Teams working exclusively on x86 or NVIDIA hardware will find the x86 path functional but not the primary target; the Mac Preview is a developer convenience, not a performance deployment. Before deploying to edge hardware, read the build guide and deployment guide in the project documentation, and verify that your target chip model (BM1688, CV186X, RK3576, or RV1126B) matches the tested release tier for v1.1. ## FAQ ### Which edge AI hardware platforms does CosmoEdge support? CosmoEdge v1.1 supports Sophon BM1688 and CV186X (using BMRT inference with .nn model artifacts), Rockchip RK3576 and RV1126B (using RKNN with target-specific .rknn artifacts), and x86 Linux and Windows (using ONNX Runtime). Apple Silicon macOS is a developer Preview under amd64 emulation. Sophon BM1684X is listed as Planned. ### What is Model Guard in CosmoEdge? Model Guard 2.3 is the commercial preset-model distribution protection component. It protects model distribution in Sophon Protected packages using model encryption and device provisioning tooling. The README states that Open and Protected packages expose identical application features; the difference is model encryption, not access to orchestration or event functionality. ### Can CosmoEdge run without edge hardware for development? Yes. The x86 Docker Compose path (docker-compose.x86.yml) runs the full orchestration console on Linux or Windows without edge hardware, using ONNX Runtime. The Apple Silicon macOS Preview path runs under amd64 emulation and is validated for one local-video workflow. Both paths open the console at localhost:8080. ## Sources - [cosmo-wander-ai/cosmo-edge on GitHub](https://github.com/cosmo-wander-ai/cosmo-edge) - [License: Apache-2.0](https://github.com/cosmo-wander-ai/cosmo-edge/blob/main/LICENSE) - [Project website](https://www.cosmowander.ai/) - [README](https://github.com/cosmo-wander-ai/cosmo-edge/blob/main/README.md) - [Releases](https://github.com/cosmo-wander-ai/cosmo-edge/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/cosmo-wander-ai-cosmo-edge