ENPT

Ainvest-Decision-27B

Last updated:

Open99–99ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity4.8/10

    Apache-2.0 BF16 weight merge on Qwen3.8-27B (26.1B, ~52 GB) with a pip package, CLI and Decision Index adapter. Text only. No hosted API, integrations or SLA.

  • Capability6.7/10

    Decision Index 0.4.1 (29 Oct) Full 68.23 (#10 of 247), ECE 0.042, median 99 ms, p95 1.07 s. Choice (2–255 options) and yes/no only; no score type. No calibration method published.

  • Adoption1/100

    About 81 downloads and 2 likes (31 Oct). No integrations or production users found.

Vendor claims

publicado (curadoria autônoma 31/10 BRT). Scores are calculated from scratch.

What it is

Ainvest-Decision-27b-v1.0 is a text-only decision model from Ainvest, built on Qwen3.8-27B. It scores every supplied option with a 255-way readout instead of generating an answer, using non-causal full-attention layers (the same prompt semantics as pplx-decider, which the card credits). The release is the evaluated “M02” weight merge, stored in BF16. The card credits the tasksource data for much of the gain. Source: Hugging Face.

It answers choice (2–255 options) and noul (probability of true). Requests over 16,384 tokens per question are refused, not truncated. Vision is explicitly not trained or claimed.

Board result

Board Row Result
Decision Index 0.4.1 (data 29 Oct) Ainvest-Decision-27B v1.0 Full 68.23 (#10 of 247), public 71.88; ECE 0.042; median 99 ms, p95 1.07 s

Not found on JevBench or the Artificial Analysis Decision Index on 31 Oct. The boards are never combined.

Vendor claims (UnverifiedClaim)

  • Public Decision Index 0.3 of 71.88 (raw 78.57) from the authors’ own completed public run (source). The board’s public column shows the same 71.88.
  • Local single-H200 wall time on a 100-request sample: median 141.67 ms, mean 994.69 ms. The card itself says this is not the official latency measurement.

Scores (from scratch, 31 Oct 2026)

  • Capability 6.7 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 6 × 0.30 + 6 × 0.25 + 7 × 0.25 + 8 × 0.20 = 1.80 + 1.50 + 1.75 + 1.60 = 6.65 → 6.7. Types 6: choice and yes/no, no score. Calibration 6: board ECE 0.042; the card names no method and tells you to check calibration on your own data. Latency 7: board median 99 ms, p95 1.07 s. Accuracy: 8 × 68.23 ÷ 59.22 = 9.22 → capped at 8.
  • Maturity 4.8 = availability × 0.30 + docs × 0.25 + integrations × 0.25 + support × 0.20 = 7 × 0.30 + 7 × 0.25 + 2 × 0.25 + 2 × 0.20 = 2.10 + 1.75 + 0.50 + 0.40 = 4.75 → 4.8.
  • Adoption 1 = engagement × 0.35 + ecosystem × 0.35 + production × 0.30 = 1 × 0.35 + 1 × 0.35 + 0 × 0.30 = 0.70 → 1.