Download

Get Hark 0.4.0

Installs once; later versions arrive as a signed in-app update you approve.

Pick the build for your machine. All of them are free, open source and identical in features; the GPU builds are just faster.

Windows, NVIDIA GPU

Uses CUDA for fast transcription and summaries: an 88-minute call is fully processed in about 10 minutes. Needs a GeForce GTX 10-series or newer with 6 GB+ VRAM and a recent driver; nothing else to install.

Download CUDA installer

Windows, any PC

Runs on the CPU. Works on any 64-bit Windows 10/11 machine with 8 GB of RAM and an AVX2 CPU (2013 or newer). Slower, but fully featured; Hark sizes the models to your machine so long calls still finish in reasonable time.

Download CPU installer

macOS, Apple Silicon

M1 or later, uses Metal. System-audio capture needs macOS 14.6+; older versions record the mic only.

Download for Apple Silicon

macOS, Intel

For Intel Macs on macOS 13 or later. CPU inference; the CPU tiers of models are selected automatically.

Download for Intel

After installing

  1. Open Hark. It checks your mic and system audio, then suggests model sizes for your hardware.
  2. Download the models once (from Hugging Face, checksum-verified, 1 to 7 GB depending on tier).
  3. Join a call. When the popup appears, check the window and audio choice and click Record. Or press Ctrl/Cmd+Shift+R anytime to open the same picker.
  4. Open the recording from the Library to follow captions and running notes while the call is on.

Unsigned builds

Hark's installers are not code-signed yet, so Windows SmartScreen and macOS Gatekeeper will warn on first launch. On Windows choose "More info, Run anyway"; on macOS right-click the app and choose Open. Every release lists SHA-256 checksums, and you can always build from source.