Tiny-Jev

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

• Maturity4.3/10

  Public Apache-2.0 weights, clear HF card and transformers API; day-old release with no hosted SLA and early traction.

• Capability5.8/10

  Full Choice / Score / Noul plus batched decide on a 0.6B stack with temperature calibration; all headline benches are author-reported UnverifiedClaim, and held-out domains drop hard.

• Adoption10/100

  Early HF footprint after a 21 Sep 2026 publish; still niche among open-Jev hunters, not broader market awareness.

Vendor claims

• Held-out test accuracy 95.8% / ECE 0.004 on author training-distribution split (22,110 items)[Vendor claim — not independently verified]
• SST-2 90.4 vs zero-shot reference 91.6 after fitting on 2k disjoint items (open-system-one protocol)[Vendor claim — not independently verified]
• ~20–50 ms per call on Apple M-series; a few ms on a consumer GPU[Vendor claim — not independently verified]

Tiny-Jev is an independent open decision model — Qwen3-0.6B with a LoRA merged into one safetensors file plus a marker-token decision head — not a TypeSafe product and not a method-only logit wrapper.

State + typed questions in; probability distributions out in a single forward pass with no text generation. Primitives: Choice, Score, and Noul (catalog boolean), plus a batched decide fan-out over one state. Card: Hugging Face lostargon/Tiny-Jev (Apache-2.0; trust_remote_code loads the shipped ~150-line modeling_tiny_jev.py).

Limits

• Author benches (in-distribution, public datasets, OOD domains) are self-reported — treat as UnverifiedClaim, not ModelSystem.One numbers
• Public-benchmark rows were fitted on 2k items per dataset vs a zero-shot reference; the card itself asks you to read that gap honestly
• Fully held-out domains land ~50–60% in the author table — expected for a 0.6B backbone
• No hosted SLA; local transformers load only at last check
• Not affiliated with TypeSafe; “Jev” in the name means the interface shape only

Why it is in the catalog

Trained decision weights + open inference that returns typed calibrated options (peer class of NanoJev and Kev, lighter than hosted Jev). Radar note: 2026-09-22.