Installation & system requirements
LYNX runs on edge and desktop hardware across six platforms. This page covers what you need to install and run the SDK.
Beta note: LYNX is in pre-release. The details below reflect what we anticipate at launch and may shift before general availability.
Install
pip install golynx
The wheel bundles the native inference engine. No separate runtime,
no PyTorch dependency, no manual CUDA setup. Model weights are fetched
on first use and cached locally. The import path is import lynx.
The package is not yet published. Beta participants receive install access directly. Join the beta for early access.
Supported platforms
| Platform | Architecture | Acceleration |
|---|---|---|
| Linux | x86_64, aarch64 | CUDA / TensorRT, CPU |
| macOS | Apple Silicon, Intel | CoreML (Neural Engine), CPU |
| Windows | x86_64 | CUDA, DirectML, CPU |
| iOS / iPadOS | arm64 | CoreML (Neural Engine) |
| Android | arm64 | CPU / GPU |
| NVIDIA Jetson | aarch64 | TensorRT |
Python versions
Wheels are anticipated for CPython 3.10 through 3.14. pip selects
the correct wheel automatically for your platform and Python version.
GPU requirements
LYNX is built edge-first. There is no datacenter-class GPU requirement.
| Model | Parameters | Reference latency |
|---|---|---|
| pico | sub-1M | runs on minimal hardware |
| nano | 0.73M | ~1.3ms (RTX 3060, TensorRT) |
| medium | — | benefits from more headroom |
| large | — | benefits from more headroom |
The smaller models (pico, nano) run comfortably on low-VRAM edge devices including Jetson Orin Nano and similar. Larger models benefit from additional GPU memory but still target edge-class hardware.
CPU-only inference is supported on all platforms via ONNX Runtime — no GPU required, at reduced throughput.
Verifying acceleration
To confirm the SDK is using your GPU rather than falling back to CPU:
LYNX_INFERENCE_LOG=1 python your_script.py
This logs the active execution provider chain (CoreML / CUDA / DirectML / CPU) when the inference session is built.
NVIDIA Jetson
On Jetson, models compile to TensorRT for the local hardware on first run. The compiled engine is cached, so subsequent starts are fast. JetPack 5 and JetPack 6 are supported.
Offline operation
Inference runs on-device — no per-frame network calls, no cloud dependency, and your video never leaves your hardware. Telemetry and trial verification use periodic connectivity when available; paid licenses can be configured for extended offline operation. See the licensing page for how activation works.