Skip to content

Releases: chriscoveries/jevalaya

jevalaya v0.1.0 — first public build

Choose a tag to compare

@chriscoveries chriscoveries released this 22 Sep 12:51

First public build of jevalaya — a fast, private, local decision router for Mac. One drop-in /predict endpoint; every request routes to Apple Neural Engine, embedded MLX, or Jev (TypeSafe), with a routing receipt on every call.

Install (Apple Silicon)

tar -xzf jevalaya-v0.1.0-macos-aarch64.tar.gz
cd jevalaya-v0.1.0-macos-aarch64
cp jevalaya.example.toml jevalaya.toml   # point it at your model dirs
export JEVALAYA_TOKEN=local-dev
./jevalaya serve --config jevalaya.toml

The tarball ships the binary, the example config, and the consumer handover doc. For Jev escalation set TYPESAFE_API_KEY. Source build: cargo build --workspace.

What's in it

  • POST /predict drop-in contract: {state, questions} in, {model, answers, usage, routing} out
  • Three backends: CoreML ANE, MLX (PyO3 bridge), Jev — graceful degradation to whatever's standing
  • Routing receipts + JSONL observability + /feedback verdicts
  • Demo suite in tools/demo/ (snake + burst race) plus recorded MP4s in docs/assets/

Measured on this M1 Max: 4-call ANE burst drains ~18ms p50 while MLX stair-steps under the serialized bridge — see the race video. Upstream laya-coreml reports ~8ms p50 on M3 Max.