ENPT

Decision 3.0 (d3)

Last updated:

Open7–553ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity6.5/10

    Six Apache-2.0 sizes (26B to 0.6B) on Qwen3.8/3.5, loadable in Transformers with trust_remote_code and a decision pipeline; detailed cards. Built for vLLM Semantic Router. No GGUF or /v1/systemone server documented yet. Community project, no SLA.

  • Capability7.6/10

    Choice / Yes-No / Score over text, JSON, images and video (2 fps). d3 27B is #1 on the official Decision Index 0.3 (12 Oct): 65.65 full, 64.25 public, tied group 1 of 144. Vendor latency 55 ms text, 553 ms 10-s video on an MI325X. No ECE published.

  • Adoption3/100

    About 460 downloads and 32 likes across the six repos on day one; announced by the vLLM project. Feeds vLLM Semantic Router. No named production users.

Vendor claims

Read this first

  1. Not on the board yet. The Decision Index data we read (regenerated 10 Oct, 07:54 UTC) predates the release. The public-suite numbers are vLLM-SR’s own runs with the official 0.3 kit. They count as evidence here, but they are UnverifiedClaim.
  2. The card also lists a full score of 64.7 (with private same-skill and new-domain parts). Private parts are run only by the maintainers, and the entry is not in the board data we read, so we do not use that figure.

Update, 12 Oct 2026: d3 27B is now on the official board: 65.65 full score, #1 (tie group 1 of 144); public part 64.25, same-skill 63.59, new-domain 58.64. The board’s full score is close to the card’s 64.7. Scores here are unchanged (accuracy was already 8).

Update, 13 Oct 2026: all six sizes are now on the official board (regenerated 13 Oct, 160 entries). The public parts match the vendor’s numbers exactly. d3 27B moved to 66.26 full, still #1.

Size Full (rank) Public
d3 27B 66.26 (#1) 65.45
d3-flash 9B 55.79 (#31) 57.79
d3-mini 4B 52.61 (#39) 54.90
d3-nano 2B 38.86 (#86) 42.22
d3-lite 0.8B 32.33 (#101) 36.93
d3-edge 0.6B 20.24 (#124) 28.82

Quantized community builds of d3-nano and d3-lite are circulating on X; we haven’t checked them, so they aren’t listed. Scores unchanged.

Update, 17 Oct 2026: on the board’s 16 Oct data (174 entries) d3 27B keeps 66.26 full but is now #2, in tie group 2, behind StartLux-Decision v1.1 27B (69.12, first and not tied); its public part (65.45) ranks #5. The other sizes rank lower as the board grew: d3-flash #43, d3-mini #52, d3-nano #100, d3-lite #115, d3-edge #138. The launch claim “#1 on both Jev Decision Index” matched the board of 12-13 Oct; it no longer does. Scores unchanged (accuracy 8 stays: still a top-2 Full result). > publicado (OK do Fabricio 17/10 BRT).

The six sizes

Model Params Base DI 0.3 public (vendor) Text latency (vendor, MI325X)
d3 26.09B Qwen3.8-27B 65.45 55 ms
d3-flash 8.39B Qwen3.5-9B 57.79 —
d3-mini 4.54B Qwen3.5-4B 54.90 —
d3-nano 2.21B Qwen3.5-2B 42.22 —
d3-lite 0.85B Qwen3.5-0.8B 36.93 —
d3-edge 0.59B Qwen3.5-0.8B 28.82 6.7 ms

All read text or JSON plus several images and videos per request (videos at 2 fps, up to 32 frames) and answer Choice, Yes/No and Score questions together, one forward pass per question. Load with AutoModel.from_pretrained(..., trust_remote_code=True) or the decision pipeline. The previous generation is Decision 2.0.

Score working (11 Oct 2026, scored on d3 27B)

  • Maturity = 9 × 0.30 + 7 × 0.25 + 5 × 0.25 + 4 × 0.20 = 2.70 + 1.75 + 1.25 + 0.80 = 6.50 → 6.5
  • Capability = 10 × 0.30 + 6 × 0.25 + 6 × 0.25 + 8 × 0.20 = 3.00 + 1.50 + 1.50 + 1.60 = 7.60 → 7.6
  • Adoption = 4 × 0.35 + 4 × 0.35 + 0 × 0.30 = 1.40 + 1.40 + 0 = 2.80 → 3

Accuracy 8: the vendor’s public-suite run with the official kit counts as evidence (Fabricio, 10 Oct). Latency 6: self-run only. Calibration 6: calibrated probabilities claimed, no ECE published.