kev: a tiny Jev-like you can train on a MacBook

A family of tiny Jev-like decision models — 0.5B at launch, now 0.6B, 4B and 8B on Qwen3.5 — using the same LoRA-plus-pointer-head recipe with block-causal masking so questions run in parallel; the HTTP API mirrors System One, so the official SDK works with only a base_url change.

Notes

Kev-0.5B: 0.799 held-out accuracy, ECE 0.031, ~160 ms for six-question requests on an Apple M5. Kev-8B: 79.6% out of domain versus Jev at 85.7%, per the author. Code, evals and weights public.

source-backed — Public repository, docs or live artifact. About this label

Destinations

Verification levels

Report a problem →

Also linked

• Hacker News thread ↗

Filed under

• #replica
• #local
• #qwen
• #github

In collections

• Local & replica models →