ENPT

Super Decider

Last updated:

Open250–890ms$0.03/M input

Quadrant scores

See the full quadrant

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

  • Maturity6/10

    Apache-2.0 NVFP4 weights (18 GB) plus a hosted /v1/systemone and OpenAI-compatible API on ASI:Cloud at $0.03/M input. HF card, CUDOS docs and a pinned vLLM runtime on GitHub. One 48 GB GPU. No public SLA.

  • Capability8.1/10

    JevBench API (v1.7.0, measured 21 Oct): composite 78.5, #1 of 48; Intelligence 78.9, Calibration 90.8, p50 0.25 s, $0.016 per 1,000 decisions. Choice, yes/no, score (plus span). Not on the Decision Index or AA.

  • Adoption1/100

    79 downloads and 8 likes on Hugging Face, 14 GitHub stars (26 Oct). Launch post from ASI:Cloud on 26 Oct. No named production users.

Vendor claims

News covering this model

publicado (OK do Fabricio 26/10 BRT). Scores calculated from scratch.

What it is

Super Decider is a decision model from Singularity Compute. The weights are on Hugging Face (Apache-2.0, uploaded 13 Oct 2026) and the serving runtime is on GitHub. You send a state (text, JSON or images) and typed questions. It returns one calibrated answer per question: a probability per option for choice, P(yes) for noul, and an expected level with its distribution for score. The vendor also lists a span type that quotes the input.

  • Base: google/diffusiongemma-26B-A4B-it, a discrete-diffusion mixture-of-experts model (26B total, 4B active), merged into NVIDIA’s NVFP4 checkpoint. The download is 18 GB.
  • Readout: all answer slots of a request are filled in one denoising step, so extra questions add little cost (vendor). The prompt block of each question is read in both directions; the runtime applies this through a small vLLM patch.
  • Thinking: when the model is unsure, it can write a short thought and read the answer again. A request can switch this off or force it on every question.
  • Serving: POST /v1/systemone in the Jev request shape, plus /v1/chat/completions. One 48 GB GPU is enough per the card. Hosted on ASI:Cloud at $0.03 per million input tokens, output free (CUDOS docs).

The runtime’s environment flags (DJEV_BIDIR_SERVE, DJEV_BIDIR) suggest a link to the DiffusionGemma stack behind djev-spark. We have not confirmed it.

Board status

Board Status
JevBench API (v1.7.0, measured 21 Oct, second regular draw) #1 of 48, composite 78.5; Intelligence 78.9, Calibration 90.8, Speed 85.3, Cost 64.3; p50 0.25 s; $0.016 per 1,000 decisions. “Thinking” variant: #8, 74.5 (Intelligence 80.0, p50 0.89 s)
Decision Index No row (Space data of 22 Oct)
Artificial Analysis No row (14 of 17 models, unchanged)

The JevBench run was on a hosted endpoint that saw the sealed inputs (API board rules). kettei-fat still has the higher Intelligence (79.8) and the top Capability Score (86.3 vs 84.9).

Limits the vendor lists

  • Answers can change with meaning-preserving edits, such as JSON key order or option order.
  • Sometimes confidently wrong on multi-step date and money arithmetic; thinking on every question fixes part of it.
  • English-centric training data. Images go through the base model’s vision path without image fine-tuning.

Author claims (UnverifiedClaim)

  • Fine-tuned on permissively licensed and synthetic decision data; no benchmark test items used in training (source)
  • 151 options in one question and up to 64 questions per call tested (source)

Scores (from scratch, 26 Oct 2026)

  • Capability 8.1 = types 10 × 0.30 + calibration 7 × 0.25 + latency 7 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.75 + 1.75 + 1.60 = 8.10. Calibration 7: JevBench axis 90.8 (≥ 90). Latency 7: board p50 0.25 s. Accuracy: 8 × 78.9 ÷ 63.6 = 9.9, capped at 8 (one independent source; not #1 on the Decision Index).
  • Maturity 6.0 = availability 8 × 0.30 + docs 7 × 0.25 + integrations 4 × 0.25 + support 4 × 0.20 = 2.40 + 1.75 + 1.00 + 0.80 = 5.95 → 6.0.
  • Adoption 1 = engagement 2 × 0.35 + ecosystem 1 × 0.35 + production 0 × 0.30 = 1.05 → 1.