StartLux-Decision

Last updated:

Open12–102ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity5.3/10

    Open family (0.8B-27B, Instant and AutoThink variants) with a technical report (arXiv 2610.17309), an Apache-2.0 serving package with /v1/systemone, GGUF and MLX builds. Weights are CC BY-NC-4.0: commercial use needs a license from StartLux Labs. No hosted API or SLA found.

  • Capability8.3/10

    Choice / noul / score with up to 8 images. Independent: Decision Index 0.3.2 Full 69.12 (#1; Jev 60.11) and JevBench open weights Intelligence 73.2, calibration 91.9, p50 55 ms on the board's GPU. No ECE published; the authors chose recipes using the DI public subset.

  • Adoption11/100

    609 GitHub stars and about 115k downloads across 53 Hugging Face repos, mostly GGUF (27B AutoThink GGUF 13.8k), plus a technical report and community GGUF builds. No hosted API, integrations or production use found.

Vendor claims

News covering this model

Non-commercial license. StartLux-Decision weights are CC BY-NC-4.0: free for research and other non-commercial use with attribution. Commercial use requires a separate license from StartLux Labs. The inference code in startlux_decision/ is Apache-2.0, but that does not extend to the weights.

StartLux-Decision is a family of open typed-decision models from StartLux Labs: 0.8B · 2B · 4B · 9B · 27B, plus a 35B-A3B mixture of experts added on 1 October. The first five cards appeared on 29 September 2026, with GGUF builds (BF16, Q8_0, Q4_K_M) that afternoon. Each size ships an in-folder startlux_decision package that serves TypeSafe-compatible POST /v1/systemone (choice / noul / score with a probability per option), so Jev clients work unchanged. Configs show the Qwen3.5 architecture (Qwen3.5-MoE for the 35B-A3B).

This is the family page (same pattern as OpenDecider / RSI-Jev). The docs URL points at the 4B card as a practical mid-size default.

Update 17 Oct 2026: v1.1, the technical report and the official board

What shipped. On 10-11 Oct StartLux Labs published v1.1 in two variants and five sizes: Instant (0.8B, 2B, 4B, 9B, 27B; one forward pass, bidirectional prompt attention in the full-attention layers) and AutoThink (4B, 9B, 27B; one pass, and when the top probability is under a 0.6 gate the model writes a short thought and mixes both read-outs). The technical report (arXiv, 12 Oct, 38 pages) describes a three-stage training run and the RLCD² reinforcement-learning step for AutoThink. Weights stay CC BY-NC-4.0; the serving package (StartLuxLabs/StartLux-Decision, 609 stars) is Apache-2.0 and serves POST /v1/systemone. Cards add GGUF and MLX builds and take up to 8 images. The 35B-A3B and the v1 checkpoints from 29 Sep remain on the Hub.

Official boards (separate sources, not averaged).

Board StartLux-Decision v1.1 27B Jev reference
Decision Index 0.3.2 (16 Oct 19:00 UTC; marked new) Full 69.12, #1 (not tied); public 72.88 (#1); same-skills 64.82; new-domains 61.41 Jev 1.13: Full 60.11 (#18), public 57.96, same-skills 58.04, new-domains 54.98
JevBench open weights (draw v1.6.28), Instant Capability 82.6; Intelligence 73.2; Calibration 91.9; self-hosted p50 0.055 s (on-GPU); cost 6.4× Jev’s reference, outside the Jev-class cap Jev 1.13: Intelligence 63.6, calibration 90.6
JevBench open weights, AutoThink (gate 0.6) Capability 80.3; Intelligence 69.0; Calibration 91.7; p50 0.056 s; cost 10.9×

The board’s public figure (72.88) matches the card’s own run (72.83). The authors say plainly that they chose training recipes and seed averages by the DI public subset, so the public part is a model-selection number; the Full score, whose other two parts are held out, is the one to rely on, and it still leads by about 3 points over the next model. The cards’ DecisionPivot numbers (86.9 pair accuracy for AutoThink against 75.4 for Instant) are the authors’ own benchmark and UnverifiedClaim.

Speed (authors, one H200, vLLM, one request in flight). 27B Instant BF16: 52 ms at 256 tokens, 98 ms at 1,024, 358 ms at 4,096, 737 ms at 8,192 (FP8 51 / 72 / 251 / 534 ms). On an Apple M5 Pro the 27B MLX build takes 0.9-1.4 s at 256 tokens. A question that thinks (AutoThink) adds its thought time.

Author benchmarks, v1 (29 Sep; UnverifiedClaim, superseded by the board rows above)

Size DI 0.2.1 JevBench public /231 Intern avg Latency (3 q, H200)
35B-A3B (MoE) 61.55 210 92.29% 52.5 ms
27B 63.88 208 91.82% 102.3 ms
9B 58.63 201 91.08% 35.7 ms
4B 52.75 204 91.17% 26.0 ms
2B 44.19 196 88.46% 15.5 ms
0.8B 38.86 179 85.03% 12.2 ms
Jev 1.13 (from the board) 57.91 199 88.74% 64.0 ms (API server time)

Latencies are end to end over HTTP on one H200 in bf16, one request at a time, with all questions answered in one forward pass. “Intern avg” is the average over the seven suites in the Intern-Decision bundle. None of the StartLux rows were on the official Decision Index board at our 2 Oct check, and the cards publish no calibration figure. Treat every number as author-reported.

The table above is the v1 release. None of its rows were on the official board at our 2 Oct check; v1.1 is now on it (section above).

Fit / anti-fit

Fit: research and non-commercial System One work that wants a size ladder, a GGUF path and a Jev-compatible local server.

Anti-fit: any commercial product without a separate StartLux license; relying on the DI / JevBench claims before board confirmation; uses that need calibrated probabilities with a published ECE.

Why it is in the catalog

Trained decision weights, a public family and a typed /v1/systemone server in the repo. A non-commercial license does not remove catalog eligibility (same call as bev-decider and the Nimble V3 note); it lowers availability in the maturity score and is flagged at the top of this page.

Score update (17 Oct 2026, headline: v1.1 27B Instant)

Scores move because two independent boards now carry the family. The 29 Sep working (5.1 / 7.0 / 5) is superseded.

  • Maturity 5.1 → 5.3 = availability 7 × 0.30 + docs 8 × 0.25 + integrations 4 × 0.25 + support 1 × 0.20 = 2.10 + 2.00 + 1.00 + 0.20 = 5.30. Docs 7 → 8: full cards with limits, a 38-page technical report, GGUF and MLX instructions. Availability stays 7 (CC BY-NC, same as bev-decider).
  • Capability 7.0 → 8.3 = types 10 × 0.30 + calibration 7 × 0.25 + latency 7 × 0.25 + accuracy 9 × 0.20 = 3.00 + 1.75 + 1.75 + 1.80 = 8.30. Calibration 4 → 7: board calibration axis 91.9 (Jev 90.6), per-type temperature published, no ECE (Nimble precedent). Latency 7: p50 55 ms measured by the board, no p99. Accuracy 6 → 9: two independent boards agree, with the #1 Full score on the Decision Index and an Intelligence of 73.2 on JevBench (basal-omni 74.1 and Quyet-1.0-Large 73.4 are slightly higher); 9 rather than 10 because the authors tuned on the DI public subset.
  • Adoption 5 → 11 = engagement 20 × 0.35 + ecosystem 12 × 0.35 + production 0 × 0.30 = 7.00 + 4.20 + 0.00 = 11.20 → 11 (609 stars, about 115k downloads, technical report).

StartLux Prism (inference engine, 24 Oct 2026)

StartLux Labs publishes StartLux-Prism (repo since 13 Oct; covered by Pandaily on 24 Oct), an NVIDIA inference engine that serves StartLux-Decision v1 and v1.1 AutoThink (4B, 9B, 27B) straight from the released GGUF files behind /v1/systemone. Catalogued separately as a runtime.

Author figures (UnverifiedClaim): 27B v1.1 on one H200, int8, 19.0 ms for a ~256-token question against 53.1 ms for the official engine in FP8 (vLLM 0.31); up to 5.3× faster on a three-question request; int8 scores 92.94 vs 92.97 for the official engine in BF16 on v1’s seven suites; answers are bit-identical regardless of batching. Scores unchanged; the weights stay CC BY-NC-4.0.