Laya MPS

Checked, not fully reviewed. It is built for Apple Silicon and we have no Mac, so we ran it on Linux with --device cpu, which the CLI supports: its 91 tests (all pass), its server, and its CLI in each memory preset. The README’s latency and memory figures are from an M5 Pro, and we did not measure our own. No Jev calls.

Laya MPS is a small server and CLI around Laya’s English laya-typed-decisions checkpoint, pinned to a Hub revision. Its feature is memory. full keeps every weight in RAM, reduced (the default) reads embeddings from disk, and minimal also streams transformer layers from disk. The project’s claim is that all three give the same answers. On our CPU they did: one request with a choice, three yes/no questions and a score gave identical answers, to the last digit, in all three presets.

The README is clear that this is “not a drop-in Jev endpoint”. A Jev-shaped request with a score criteria list was refused with 422 until we renamed it levels, and score probabilities come back as a list, not keyed by level. The answer types and fields otherwise follow Jev’s. It runs one inference at a time and says so. Concurrent callers get a 503 with Retry-After: 1, so put a queue in front of it if more than one thing calls it.

The model’s limits come through unchanged. For “I was charged twice for invoice 4411 and want a refund”, “Is this a billing problem?” scored 0.696, “Is it false that this is a billing problem?” 0.693, and “Is this about something other than billing?” 0.580. Phrase conditions positively, as our Laya review found.

MIT, 22 stars.

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

Review updated Oct 7, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.