Two Jevs, One Dungeon Crawl (Altered Craft)
External post by Sam Keen: running Open Jev locally on a MacBook with MLX, then pointing a Godot text adventure built for the TypeSafe Jev API at both models and comparing answers and confidence.
What this guide points to
Sam Keen’s Altered Craft post (2026-10-06) has two parts. First, a setup walkthrough for running Open Jev (the MLX 4-bit build) on Apple silicon, served through its helper shim on the same /v1/systemone request shape as TypeSafe’s API. Second, a comparison: The Sunken Crypt, a small Godot text adventure where each turn’s possible moves become the options of one Choice question (plus “none of these”), is pointed at both TypeSafe’s jev-1.13.0 and the local Open Jev by switching an environment variable.
On the author’s 16-line test set, both models got every line right and gave the same answer on each of three runs. They differed on confidence: TypeSafe’s model gave 0.81 to 0.89 on the slang lines, Open Jev 0.96 to 0.99, and Open Jev’s values did not change between runs. The author does not draw conclusions about latency because the laptop was busy.
Treat it as a community experiment, not a TypeSafe product page. Open Jev weights are CC BY-NC 4.0 according to the post; check the license before commercial use.
This page is an unofficial field manual pointer — confirm SDK names, limits, and pricing against current docs.
What to take away
- A game’s parser can become one Choice question per turn, with confidence bands for act (0.75 or more), ask which move was meant (0.40 or more), or reject.
- Two backends that share the request shape can be swapped with configuration, which makes side-by-side checks cheap.
- Thresholds are model-specific: with a 0.9 bar, the two models would have behaved differently on four lines.
Source
Click through: Two Jev’s One Dungeon Crawl (Altered Craft, Sam Keen)