ENPT

Decision Fast

Last updated:

Open100–300ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity3.8/10

    Preview release with frozen eval weights; open LoRA + head + inference code on HF, but EVALUATION-PERMISSION only — no general commercial licence asserted.

  • Capability6.2/10

    Official JevBench (17 Oct): Intelligence 6.1, calibration 75.1, p50 0.058 s.

  • Adoption9/100

    Brand-new preview; niche interest among cost-sensitive builders — no production footprint yet.

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 Fast is an ultra-cheap 0.6B decision model from FlyMy.AI — LoRA adapter + pointer head on Qwen3-0.6B-Base (checkpoint qwen3_06b_headfirst_ep2a_v53, 606M params / ~10.6M trained) — not a TypeSafe product and not a production release.

One forward pass: state + typed question (choice, score, noul → catalog boolean) in, calibrated probability distribution out — no output tokens, no decode loop. Weights: flymy-ai/decision-fast-preview. The repo includes the LoRA adapter, structural-token rows, pointer head, and executable inference code.

Why it matters: ultra-low cost

At ~$0.0063 per 1,000 decisions (JevBench v1.4.2 cost metric: 76.1), Decision Fast ties Kev 0.5B/0.6B as the cheapest decision model in the benchmark. For high-volume, cost-sensitive workloads where calibration quality is secondary, this is the entry point.

JevBench v1.4.2 Score
Overall 32.5 (#33)
Intelligence 37.1
Calibration 65.3
Speed 81.6
Cost 76.1

Trade-off is clear: you get ~$0.006/1k decisions, but the Hard-task accuracy drops to 39.64% and calibration lags behind larger models. For volume routing, soft labels in pipelines, or pre-filtering before a heavier model, that may be the right trade.

Limits (read before you ship)

  • Preview / evaluation only — EVALUATION-PERMISSION.md grants benchmark runs; a general commercial weight licence is not asserted.
  • Backbone is 0.6B; JevBench #33 overall means this is not a frontier model — transfer to hard OOD tasks will be weak.
  • Calibration score (65.3) is lower than most peers; fit temperature on your labels if confidence matters.
  • Training mixture includes ~96 SNI/Gigaword-derived records with unresolved upstream entitlement; see provenance-review.json.
  • 4096-token packing limit — long state text may be truncated.
  • Sibling of Decision 2B (not yet in catalog); Fast uses the earlier v7 mixture, not the H8/H9 Fable additions.

Run locally

python -m venv .venv && . .venv/bin/activate
pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cu128
pip install numpy==2.1.2 -r requirements-minicpm.txt
python model.py --verify-only
import model
decider = model.load()  # once per process
result = decider.decide(
    "The customer was charged twice.",
    {
        "type": "choice",
        "instructions": "Choose the support queue.",
        "criteria": {"billing": "Billing and payments", "technical": "Technical support"}
    }
)
print(result["probabilities"])

JevBench harness integration:

git clone https://github.com/fstandhartinger/jevbench.git
git -C jevbench checkout 218511d85c5e9fc3dd12a08fae0e9818ea48bb68
python run_jevbench.py --harness ./jevbench --output ./new-results --price-input-per-million 0.01

Why it is in the catalog

Public LoRA + pointer-head weights with typed decision I/O, documented inference code, and the lowest cost-per-decision tier alongside Kev. Preview status and weaker calibration keep it below Decider, NanoJev, and Kev on maturity/capability — but for pure cost optimisation on high-volume workloads, it is the starting point.

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
JevBench open weights (live 17 Oct) Decision Fast Capability 40.6; Intelligence 6.1; calibration 75.1; p50 0.058 s; $0.0046/1k

Board-status statements on this page to recheck against the table (may be superseded):

  • “At ~$0.0063 per 1,000 decisions (JevBench v1.4.2 cost metric: 76.1), Decision Fast ties Kev 0.5B/0.6B as the cheapest decision model in the benchmark.”
  • “Backbone is 0.6B; JevBench #33 overall means this is not a frontier model — transfer to hard OOD tasks will be weak.”

Score update (audit, 17 Oct 2026). Capability 5.0 → 6.2, first full working (the earlier score had no stored breakdown) = types 10 × 0.30 + calibration 5 × 0.25 + latency 7 × 0.25 + accuracy 1 × 0.20 = 6.20 → 6.2. Types from decisionTypes; calibration from the JevBench axis 75.1 (≥90 → 7, 85–89.9 → 6, 75–84.9 → 5, 60–74.9 → 4, below 60 → 3; no author ECE counted); latency from the board’s self-hosted p50 0.058 s without p99 (≤0.25 s → 7, ≤0.5 s → 6, ≤1 s → 5, above → 3); accuracy JevBench Intelligence 6.1 ÷ 63.6 × 8 = 0.8, half-up. Maturity and adoption unchanged.

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