The April 2026 open source AI CCTV ecosystem split into two architectures. Most lists do not say which.

The April 2026 lineup splits into two camps. Classifier-on-live-path projects: Frigate (the most popular self-hosted AI NVR, MobileDet on Coral TPU by default, YOLO via ONNX or TensorRT, integrated with Home Assistant), Viseron (Python, configurable detector backends including CodeProject.AI and DeepStack), Shinobi (Node.js NVR with TensorFlow plugin pipeline), Kerberos.io (containerized agent + machinery + hub with on-agent ML), and Tokay Lite (an ESP32-based AI camera dev board for very small deployments). VLM-on-retrieval-path projects: DeepCamera (SharpAI, January 2026 onward heavily VLM-centered with Qwen, DeepSeek, SmolVLM, LLaVA, plus YOLO26 for lighter triggers, agentic via Telegram, Discord, Slack). ZoneMinder still exists, still works, still uses pixel-diff with optional ML plugins, but it is not really an AI-first project. The split below is what determines which one fits a given site.

Trigger path = the code that decides 'fire an alert right now.' This has to run inside the frame budget of the camera feed, which on a 30 fps feed is about 33 milliseconds per frame. If the model on this path takes longer than the frame budget, you start dropping frames, and dropped frames during a fast event mean the event is missed. Retrieval path = the code that answers 'find me the clip where the person in the red coat walks past the mailroom.' This is asynchronous, runs against stored footage or stored thumbnails, and is allowed to take seconds per call because no one is waiting on the door. The two paths want different models. The trigger path wants a small quantized classifier (YOLO-nano-class, MobileNetV3, EfficientDet-D0 at int8). The retrieval path wants a VLM with broad semantics (Qwen-VL, SmolVLM, LLaVA). Most April 2026 roundups conflate them by listing 'AI CCTV' as a single capability, so a property operator who picks DeepCamera for the live path is silently mis-architected.

Because the event window is roughly 2 seconds. A tailgater enters the frame, the door is already swinging shut, and either the system fires an alert before the door latches or the unauthorized person is inside the secured space. At 30 fps that is about 60 frames. To detect the entry and emit a routable alert in under 60 frames you need a forward pass that is much smaller than the frame budget so there is room for downstream zone, dwell, and dedupe logic. A quantized YOLOv8-nano runs in roughly 12 milliseconds per forward pass on a small NPU, which leaves about 21 milliseconds of headroom inside the 33 millisecond slot. A SmolVLM-256M call on the same hardware tier needs hundreds of milliseconds per image. The VLM finishes its first inference after the door has already latched, and its second inference after the tailgater is already on the stairwell. That is the architectural force, not a benchmark preference.

Not wrong. Mis-applied. DeepCamera in 2026 is the strongest open source agentic-VLM project for video. It is built so that you can ask 'what happened at my front door this morning' in natural language and have an LLM-driven agent reason over recent footage, remember context across days, and chat back through Telegram, Discord, or Slack. That is the right tool for retrieval and post-hoc investigation. It is the wrong tool for the live trigger on a tailgate, a package theft in progress, or a parking lot loiter ramp because the model latency is wrong by an order of magnitude. The honest deployment is DeepCamera (or any VLM-on-retrieval project) running alongside a fast classifier-on-live-path project, not in place of it.

Frigate's default detector is MobileDet SSD running on a Google Coral TPU at low single-digit milliseconds per inference, well inside the frame budget. Frigate also supports YOLO variants via ONNX, TensorRT, OpenVINO, and HailoRT for users with a small GPU or a Hailo accelerator. The relevant property is that the live path is bounded to a small fast classifier; the slow stuff (clip indexing, semantic search) sits in separate processes. This is the architecture every classifier-on-live-path project converges on. The differences between Frigate, Viseron, Shinobi, and Kerberos.io are in plumbing and ergonomics, not in the architectural choice.

On Cyrano specifically, the live path is one quantized classifier doing one forward pass over the DVR's 1920x1080 HDMI composite frame, then mapping the bounding boxes back to per-camera tile coordinates. The active model on the unit is yolov8n-int8-psv-v4.2.1, person/vehicle/package classes, 12.1 milliseconds per forward pass on the on-device NPU, with a MobileNetV3-small classifier head as a fallback for low-light tiles. One forward pass covers all 16 to 25 cameras because the DVR has already arranged them as tiles. After the box, the pipeline runs a per-tile zone filter, a per-zone dwell filter, an arming-window check, then appends to the on-device outbox NDJSON file. None of that pipeline involves a VLM. The VLM, when it runs, runs offline against thumbnails for retrospective natural-language search.

Yes, with caveats. Frigate is a great fit if your property's cameras are IP cameras that expose RTSP per camera, you have a server you maintain, and you have someone in-house who is comfortable tuning detector configs. That setup is common in newer Class A buildings and rare in Class B and C buildings where the existing camera plant is analog or HD-TVI on coax wired to a DVR. DeepCamera is a great fit when you want a chat-with-your-cameras agent and you can tolerate retrospective rather than real-time alerting. Most multifamily operators want both: a fast trigger so the on-call manager gets a phone call, and a retrieval layer for incident review. Cyrano ships that pair as a single device that plugs into the DVR HDMI; Frigate plus DeepCamera ships as two open source projects you stitch together.

Frigate is most often deployed on a small x86 mini-PC or NUC with a Google Coral USB or M.2 TPU, or on a Jetson Orin Nano. Viseron and Shinobi run on similar hardware and increasingly on a Jetson or a Hailo-8L PCIe card. Kerberos.io is container-first and runs anywhere Docker runs, with optional GPU. Tokay Lite is the smallest of the bunch, ESP32-S3 plus an OV-class image sensor, intended for one or two camera deployments, not 16-25 channel sites. DeepCamera in 2026 targets Mac Mini and AI PC class machines because the local VLMs need 8 to 16 GB of memory and either Apple Silicon unified memory, a small dedicated GPU, or an NPU rated for large-model inference. The hardware ladder maps to the architecture: the smaller the model, the smaller the box; VLMs push the box class up.

ZoneMinder still works and is still maintained at a slow cadence. Its default analytics layer is pixel-diff motion detection from before the deep learning era, with optional plugins for YOLO via a Python helper. It is fine if you want a pure NVR with light AI and you accept the ergonomics of an early-2010s Perl plus PHP project. MotionEye is largely unmaintained as of 2026. iSpy / AgentDVR is only partially open source and the AI features sit behind a paid SaaS tier. None of these projects sit cleanly in either the classifier-on-live-path or the VLM-on-retrieval-path camp; they are the previous generation. They are worth knowing about for context but they should not be the architecture you choose in 2026.

Two questions. First, what is your camera input layer? If it is per-camera RTSP or ONVIF, the open source category is available and Frigate plus DeepCamera is a credible stack you can self-host. If it is a DVR with cameras over coax (common in Class B/C multifamily, gas stations, small commercial), the open source category is mostly unavailable because none of those projects ingest a DVR HDMI composite at production reliability. Second, what is your engineering tolerance? Frigate plus DeepCamera at 16 cameras is roughly 60-90 hours of standup, plus per-property tuning, plus an ongoing operations burden when models update or a Coral TPU gets disconnected. Cyrano is a $450 unit that auto-detects the DVR over HDMI EDID, ships the live classifier and the retrieval layer pre-wired, and posts to WhatsApp and SMS. Open source is the right answer when the engineering time is free; a flat-priced unit is the right answer when it is not.

Eventually, partly. The 2026 trajectory is real: small VLMs (SmolVLM, MobileVLM, MiniCPM-V) are getting faster, and quantized inference on consumer NPUs is closing the gap. But the frame budget on the live trigger path is fixed by the camera feed (33 ms at 30 fps, 50 ms at 20 fps), and VLM inference is doing fundamentally more work per frame because the model is general-purpose. A purpose-trained nano-scale classifier will stay an order of magnitude faster than a general-purpose VLM at the same hardware tier for years. The realistic 2027-2028 picture is small VLMs taking over the dwell and the zone classification (still after the trigger), while the trigger itself remains a small classifier. The split moves; it does not collapse.