NanoJev

Quadrant scores

See the full quadrant

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

• Maturity5.5/10

  Public weights, training pipeline, and game demos; independent research posture without a hosted SLA.

• Capability6.5/10

  Choice / score / boolean on a 0.6B backbone with recorded maze/Snake evals; head-to-heads with Jev are author-reported, not ModelSystem.One benches.

• Adoption10/100

  Visible in the open-Jev GitHub wave; known mainly to people already hunting replicas, not the broader market.

Vendor claims

• Local safety model ~77.8% test / ~76.6% OOD on author game questions[Vendor claim — not independently verified]
• Maze/Snake side-by-side demos vs hosted Jev (author recordings)[Vendor claim — not independently verified]

NanoJev is an independent open decision model — Qwen3-0.6B with structured decision heads — not a TypeSafe product and not a LoRA-only adapter.

State + typed questions in; full probability distributions out with no output-token decode. Supports dynamic Choice (2–255), Boolean, and ordered Score (2–10). Weights and data: Hugging Face C-Tianyu/NanoJev · source TianyuCodings/NanoJev.

Limits

• Research / game-controller demos (maze, Snake); not a drop-in hosted SLA
• Author comparisons to Jev are self-reported recordings and frozen suites — not ModelSystem.One benchmarks
• Small backbone (0.6B): expect weaker transfer than frontier hosted Jev on open semantic tasks
• Latency depends on your GPU/CPU; catalog range is a placeholder until we measure

Why it is in the catalog

Public typed decision I/O, open weights + training pipeline, and a complete backbone+heads stack (peer to Laya; heavier than Kev LoRA). Roundup context: @studio_yebisu.