CosmoEdge targets a gap that most inference stacks leave open: getting a video model off a workstation and onto the specific NPU box a customer already bought. It is a C++17 edge AI engine for video analytics and on-device VLM, published under Apache-2.0 by the cosmo-wander-ai organization. Since going public on 9 June 2026 it has reached 577 stars and 150 forks, and the current release is v1.1.0, published on 24 August 2026 — the third release in the project's history.

The organizing idea is one engine architecture with per-target inference backends. A single build selects one backend and runs models compiled for that silicon, while the ingest, orchestration, event, and observability workflow stays identical across platforms. The v1.1 release covers Sophon BM1688 and CV186X (BMRT runtime, .nn artifacts), Rockchip RK3576 and RV1126B (RKNN runtime, target-specific .rknn, with RKLLM v1.3.0 packaged only for RK3576), and x86 Linux and Windows via ONNX Runtime. Apple Silicon macOS is explicitly a Preview path running under linux/amd64 emulation for one local-video developer workflow — the README is careful to say this is not native performance evidence. BM1684X is listed as planned, not shipped.

The VLM support is the part that makes this multimodal rather than a plain detector runtime. Prompt-driven vision runs alongside conventional CV pipelines: VLM state judgment and GroundingDINO open-vocabulary detection, with an optional VLM review step gating whether a detection alarm is actually reported — a practical way to cut false positives without a second full pipeline. The published benchmark pack validates VLM at 0.1 FPS per channel across 6 channels on BM1688 and CV186X, and 4 channels on RK3576. RV1126B VLM is stated to be outside the v1.1 release claim. Those are deliberately narrow, protocol-bound numbers rather than headline throughput figures.

Below the model layer, the Rockchip data path is where the engineering shows: targeted DMA-BUF-to-RGA input, persistent bound-input, native quantized output, and direct YOLOv8 tensor decoding with explicit fallbacks. One digest-pinned Rockchip builder shares the compiler and RKNN SDK while selecting an isolated MPP/RGA profile per chip, and every build exports target markers and SHA256SUMS under build_output/. Around that sits an application layer most serving projects skip entirely — browser-based visual pipeline orchestration, model import and validation, zones, counters, OSD, alarm snapshots, event history, and structured delivery over REST, WebSocket, MQTT, or HTTP webhooks.

The caveats are real and mostly about boundaries. The engine and console are Apache-2.0, but certified hardware, commercial preset models, and Model Guard 2.3 distribution protection sit outside that line — the project states Open and Protected packages expose the same application features and differ only in model encryption and device-provisioning tooling, but teams should read that boundary before committing. The repository is also young, with 8 contributors and 7 open issues, and the benchmark evidence is scoped to a recorded protocol and a controlled 72-hour dual-CV profile rather than broad field data. For teams deploying video AI onto Sophon or Rockchip hardware, it is an unusually complete starting point; for anyone on other silicon, the platform list is the first thing to check.