ENPT

Decision 2B

Last updated:

Open80–200ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity4.5/10

    Fully documented eval package with reproducible harness, but preview status and EVALUATION-PERMISSION license restrict commercial use; not a production release.

  • Capability5.8/10

    Official boards (22 Oct): Decision Index 0.4.1 Full 39.80 (#122) for FlyMy's Decision 4B v1.6 (the board has no 2B row); JevBench Intelligence 22.9, calibration 86.8, p50 0.075 s.

  • Adoption12/100

    Early in the open-jev wave; evaluation-only license and preview status limit adoption to benchmark researchers and model builders.

Vendor claims

publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

Decision 2B is an evaluation preview by FlyMy.AI (@denti) — LoRA adapter + pointer head on frozen MiniCPM5-2B — not a TypeSafe product, not a production release, and not the same as Decider-2B (Mapika’s Qwen3.5-2B full fine-tune).

It returns native decision probabilities (choice, score, noul → catalog boolean) via pointer-head softmax — no generated tokens, no logprob parsing. Checkpoint: minicpm5_reduced_v16_4k_v59. Weights: flymy-ai/decision-2b-preview. Sibling: Decision Fast (0.6B small-model track).

JevBench v1.4.2 Results

Metric Value
Overall Score 35.8 (#25)
Intelligence 38.8
Calibration 74.1
Speed 84.3
Cost 62.5 (~$0.018/1k)

Public subset measured by the author (not official JevBench rank): 75.32% overall — Easy 100%, Standard 83.33%, Hard 59.46%.

Limits (read before you ship)

  • Evaluation preview only — EVALUATION-PERMISSION.md permits benchmark evaluation; general commercial weight license is not asserted.
  • Author-measured benches (public JevBench subset) are self-reported — not ModelSystem.One evals; comparison claims are UnverifiedClaim.
  • Provenance notes unresolved: 96 Gigaword-derived records, 543 claude-fable-5-1 synthetic records with unclear upstream entitlement.
  • 4096-token packing limit; long state text may truncate. Max 255 options per request.
  • Do not put this on a production shortlist without EVALUATION-PERMISSION review and your own labelled eval.

Not Decider-2B

This model is distinct from Decider-2B by Mapika:

Decision 2B (FlyMy.AI) Decider-2B (Mapika)
Base MiniCPM5-2B Qwen3.5-2B
Method LoRA + pointer head Full fine-tune + RL calib
Status Evaluation preview Open release
License EVALUATION-PERMISSION Standard open

Why it is in the catalog

Public pointer-head decision architecture with documented JevBench reproduction, explicit provenance disclosure, and memory/latency measurements. Demonstrates MiniCPM5 as a viable decision-model backbone — useful for builders exploring non-Qwen architectures.

FlyMy’s newer release: Decision 4B v1.6 (11 Oct 2026)

flymy-ai/decision-4b-v1.6 (“FlyMyJev-4B”) is a LoRA (rank 16) on Qwen3.5-4B that reads the option letters’ logits, with temperatures per question type. Unlike the 2B preview, it is Apache-2.0. It takes up to 676 options and 131,072-token inputs and ships a /v1/systemone server that JevBench’s and the Decision Index kit’s adapters can drive. FlyMy’s numbers (UnverifiedClaim): JevBench public hard 76.6 (85/111), standard 100, easy 100, i.e. 205/231 overall, same as v1.5; p50 18.3 ms per short decision on an RTX 4090. Not on any board. This page’s scores are for the 2B preview and are unchanged; v1.6 is a candidate for its own page.

Board audit (17 Oct 2026)

publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board Row Result
Decision Index 0.3.2 (16 Oct 19:00 UTC) FlyMy.AI Decision 4B v1.6 Full 43.65 (#85, tie group 20); public 42.7 (#88); same-skills 43.09; new-domains 40.49
JevBench open weights (live 17 Oct) Decision 2B Capability 54.8; Intelligence 22.9; calibration 86.8; p50 0.075 s; $0.013/1k
JevBench open weights (live 17 Oct) Decision 4B v1.2 Capability 60.4; Intelligence 32.4; calibration 88.4; p50 0.019 s; $0.017/1k
JevBench open weights (live 17 Oct) Decision 4B v1.1 Capability 57.2; Intelligence 29.1; calibration 85.2; p50 0.019 s; $0.017/1k
JevBench open weights (live 17 Oct) Decision-4B Capability 59.1; Intelligence 37.9; calibration 80.3; p50 0.047 s; $0.014/1k

Score update (audit, 17 Oct 2026). Capability 5.8 → 6.0 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 6.8 × 0.30 + 7 × 0.25 + 4 × 0.25 + 6 × 0.20 = 5.99 → 6.0. Types, calibration and latency come from the earlier working (6.8 / 7 / 4); only accuracy is re-derived from the board (6; was 5): Decision Index Full 43.65 ÷ 60.11 × 8 = 5.8, half-up. Maturity and adoption unchanged.

Hugging Face on 17 Oct (flymy-ai/decision-2b-preview): 0 downloads, 0 likes, licence tag none, last modified 2026-09-23.

Board audit (20 Oct 2026, DI 0.4)

publicado (OK do Fabricio 20/10 BRT).

Decision Index 0.4 (20 Oct; Full now from the private parts only) lists FlyMy.AI Decision 4B v1.6 at Full 40.24 (#105); it was 43.65 on the archived 0.3.2 edition. Accuracy rule, reference Jev 1.13 on DI 0.4 Full (61.26): 8 × 40.24 ÷ 61.26 = 5.25 → 5 (was 6). Capability = 6.8 × 0.30 + 7 × 0.25 + 4 × 0.25 + 5 × 0.20 = 5.79 → 5.8 (was 6.0). Other sub-scores, maturity and adoption unchanged.

Board audit (22 Oct 2026, DI 0.4.1)

publicado (OK do Fabricio 22/10 BRT).

Decision Index 0.4.1 (22 Oct; reference Jev 1.13 Full 59.22, was 61.26 on 0.4) lists this page’s row at Full 39.80 (#122). Accuracy rule: 8 × 39.80 ÷ 59.22 = 5.38 → 5 (was 5). Capability unchanged. Maturity and adoption unchanged.