jevmlx

What jevmlx does

jevmlx brings the Jev pattern — finite fields, logits scoring, JSON assembled in code — to local MLX models on M-series Macs. You declare a schema of booleans, enums, and multi-selects (Pydantic models work well), pass a context string, and receive a validated object plus per-field probabilities from a single batched prefill. There is no free-text generation step for the decision itself: each field’s allowed options are scored from logits and the winners become the payload. The CLI ships presets such as support triage, a local HTTP server, and model aliases (quality / fast / test) that resolve to mlx-community Qwen instruct checkpoints. Design guidance in the project stresses mutually exclusive options, escape hatches when the list may be incomplete, ordered enums for severity scales, and abstention thresholds fitted on labeled data. Choose jevmlx when Apple Silicon is your runtime and you want offline or low-latency typed decisions without depending on the TypeSafe hosted endpoint.

Where it fits in the Jev stack

• Local support-ticket triage on a MacBook without cloud round-trips
• Batch classification jobs over files already on disk
• Prototyping Choice / Score / Noul-style schemas before wiring a hosted key

Browse more in Benchmarks & research, or return to the Jev AI tools directory.