OpenDMG
kevMac icon

kevMac for Mac

Paste a document, ask questions as simple forms, and get calibrated answers with probability bars.

By arinltte. Free and open source, MIT licence.

Get it with OpenDMG View source on GitHub

kevMac at a glance

Version
0.3.0
Released
4 October 2026
Download size
1.2 MB
Works on
Apple silicon and Intel
Licence
MIT
Category
Artificial Intelligence
Developer
arinltte
GitHub stars
51

About kevMac

Questions in. Answers out.

How to install kevMac on a Mac

  1. Get OpenDMG. It is a free app store for open source Mac apps. Download OpenDMG.
  2. Search for kevMac and press Get. OpenDMG downloads version 0.3.0 from arinltte's own GitHub release.
  3. Open it. The download is checked against its SHA-256 fingerprint and its signature is checked on your Mac before it lands in Applications.

Updates to kevMac show up in OpenDMG and install with one button. No account is needed.

Prefer to do it by hand? The developer's file is on GitHub: kevMac-macOS-v030.dmg (1.2 MB).

What is new in kevMac 0.3.0

# kevMac v0.3.0

The kev 1.0 release: the MLX backend arrives on Apple Silicon, every served checkpoint is pinned, documents run 8× longer with honest limits, and the app can finally see — and clean up — what it has downloaded.

## What's new

- **The kev 1.0 engine, pinned.** The engine is now fetched from the `kev-1.0` release tarball instead of a moving `main` snapshot, and stamped (`.engine-tag`) so the app knows exactly what it is running. Existing installs detect a stale pre-1.0 snapshot on launch and update it automatically — preserving the Python environment and downloaded weights — then re-sync the dependencies so **mlx-lm** arrives. That update is the whole point: the Qwen3.5 hybrid checkpoints now serve through **Metal (MLX)** instead of crawling on slow reference kernels, and a document is no longer silently truncated at 8,192 tokens.
- **Pinned weights, `@v1.0`.** Every served checkpoint is pinned to the upstream Hub tag (`--run jaredpalmer/kev-0.8b@v1.0`), and downloads fetch the exact same revision, so `HF_HUB_OFFLINE=1` resolution works and a model can never silently change underneath you (upstream moved kev-9b's `main` to new v2 weights on 2026-09-30; the pin makes that irrelevant). The pin is passed only once the engine stamp says it understands `@revision`.
…

More Artificial Intelligence apps

All Artificial Intelligence apps