OrcaCyber-Nano-1.0 is a lightweight model released under the OrcaCyber project that can be executed entirely on your own device using the Stratat desktop client. It is designed to bring small-scale cyber threat analysis to local machines without sending sensitive information to cloud services. The model is supported directly by the Strata engine, which compiles and runs the model on your CPU/GPU mix.

Why a local model like OrcaCyber-Nano-1.0 exists

The main reason to run a cyber security model on-device is data privacy: your logs, network captures, and user inputs stay on your machine. Running locally also makes the system resilient to internet outages and avoids sending files to third-party vendors. OrcaCyber-Nano-1.0 targets users who want a useful detection model without needing a datacenter.

How OrcaCyber-Nano-1.0 works with Stratat

The Stratat desktop app acts as the launcher and runner for the Strata engine, which loads the model weights and presents results on your screen. You install Stratat, load the OrcaCyber-Nano-1.0 model package, and then feed it small samples of network traffic or file metadata. The engine reports predictions and resource usage in real time, so you can see exactly how many megabytes of memory, seconds of CPU time, and milliwatt usage the model takes.

MetricValue (typical)
Model size (disk)about 1.2 GB
RAM needed for inferenceabout 1.6 GB
CPU time per sampleabout 2.5 seconds
Estimated power draw during inferenceabout 22 W

Power draw is derived from a typical laptop CPU running the inference task; you can verify it on your own machine using the OS power meter or a plug meter while running the Strata engine with OrcaCyber-Nano-1.0 loaded.

Core definition: The term "on-device inference" means that the model runs completely on the local hardware of your computer, so no data leaves the machine during processing.

Measuring the model on your own machine

To check whether the model fits your hardware, open Stratat and load OrcaCyber-Nano-1.0. Run a standard test batch of small network log snippets through the engine, then record three numbers: (1) disk space occupied by the model package, (2) peak RAM used while a sample is being processed, and (3) wall-clock time for each inference step. You can also monitor your laptop's power consumption while the engine is running; on common consumer laptops the inference phase typically draws between 15 W and 30 W depending on the processor generation.

If you see the model fitting under 2 GB of free RAM and completing a sample in roughly 2 to 3 seconds, you have confirmed that the Nano-1.0 variant fits comfortably on mid-range CPUs with mixed GPU support via the Strata engine.

What to expect in practice

The model is intentionally small so that it can run on older laptops and desktops without special cooling. Prediction quality for short network samples is modest but consistent, and because everything stays local, you can use it offline in any environment that meets the minimum hardware thresholds. This makes it suitable for personal security testing, home network monitoring, and learning how AI can help spot anomalies without uploading data to external services.