Tiny-Jev — open 0.6B typed decisions with Choice / Score / Noul
On 21 September 2026, independent author lostargon published Tiny-Jev on Hugging Face: an open ~596M System One–style decision model that takes a shared state plus typed questions and returns calibrated distributions — no autoregressive text output.
The card documents Choice, Score, Noul, and a batched decide helper on a Qwen3-0.6B stack with a LoRA merged into one safetensors file (Apache-2.0). Loading uses transformers with trust_remote_code for the shipped modeling file. The author explicitly disclaims any TypeSafe affiliation.
UnverifiedClaim / hygiene: held-out accuracies (e.g. 95.8% in-distribution), public-benchmark rows (SST-2, AG News, Emotion, BANKING77…), and the ~20–50 ms M-series latency band all come from the HF card. Public-bench rows note a 2k-item fit versus a zero-shot reference — not ModelSystem.One measurements.
Catalog entry: /models/tiny-jev. Peer context: NanoJev, Kev, hosted Jev.