SNAP

Checked, not fully reviewed. We ran the v0.5.0 release binary, verified against its published checksum, let it fetch its 1.5 GB default model, and drove it with TypeSafe’s own JS SDK. No Jev calls.

Of the three servers we drove with the official SDK today — NotJev, sys1 and snap — this is the only one that worked unchanged. systemOne() with the model set to snap1-2b or left as jev-latest returned answers keyed by question name, each with status and coverage beside the probabilities. Identical requests repeated to the digit, including the first ask of a state it had never seen — with one exception: the first request after the server started scored a Score question 0.6714 where every repeat said 0.6934. A score also moved when other questions shared its request (0.6794 alone), which is the cost of the shared prefill the design is built on.

The README’s speed figures are from an M1 Max with Metal. On four CPU cores, one question took 788 ms and twenty took 11.2 s, about 550 ms each after the first. The process held 2.9 GB.

On our negation probe the default model sat near 0.5–0.6 for “Is it false that this is a billing problem?” on billing messages, so phrase conditions positively. Outbound traffic is the model download and a daily release check to api.github.com, which prints on stderr only and is switched off by SNAP_NO_UPDATE_CHECK. MIT, 22 stars.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Oct 4, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.