ENPT

Turion Jev 31B

Last updated:

Open116–352ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity5.7/10

    Apache-2.0 BF16 weights (LoRA merged into gemma-4-31B-it), also in the Turion platform catalog. Detailed card, Decision Index kit engine and a bundled /v1/systemone server. vLLM; community GGUF. No SLA.

  • Capability7.1/10

    Choice, yes/no and 0–5 score, up to 255 options. Per-type temperature, no published ECE (board measures 0.019). Official Decision Index 0.4.1 Full 70.47 (#4 of 243; Jev 1.13 59.22). Median 144 ms, p95 815 ms on the board.

  • Adoption1/100

    24 downloads and 2 likes on Hugging Face three days after release (25 Oct); a community GGUF appeared on 25 Oct. No integrations or production users found.

Vendor claims

publicado (OK do Fabricio 25/10 BRT). Scores are calculated from scratch.

What it is

Turion Jev 31B (Hugging Face, Apache-2.0, published 22 Oct 2026) is a decision model from Turion AI, fine-tuned from google/gemma-4-31B-it on Turion’s own training platform. You send a state and typed questions (multiple choice, yes/no, a 0–5 score). It reads one probability per option from a single forward pass and generates no text.

  • Readout: each option gets a one-token code (A–Z, then two-letter codes that are single Gemma tokens), up to 255 options per question. The answer comes from the next-token log-probabilities of those codes, normalised over the options shown.
  • Calibration: one temperature per question type (choice 1.18, yes/no 1.18, score 1.12), fitted on held-out data. No ECE is published.
  • Serving: BF16 weights; the reported run quantises online to FP8 on vLLM. A TypeSafe-compatible POST /v1/systemone server ships in the repo, though it takes fewer options than the benchmark engine.
  • Limit: prompts over 65,536 tokens are refused, never truncated. Attention is Gemma’s causal attention.

Training used about 157k rows: the commercial-use subset of tasksource/tasksource-jev-typed-decisions, about 24.5k synthetic tasks labelled by Gemma-4-31B-it (the most uncertain 12% relabelled by DeepSeek V4.1 Flash), SargeDev/jev-distill-corpus-v3 and annotator soft labels. Turion says it checked every row against the Decision Index 0.3 suite and dropped 4,395 overlaps.

Board status

Board Status
Decision Index On the 0.4.1 board since 28 Oct: Full 70.47, #4 of 243 (see “Official board result” below). Before that, submitted on 24 Oct (PR #286) with only Turion’s own run of the public 0.3 suite: 66.16
JevBench, Artificial Analysis No row

Turion’s homepage lists it fourth on the Decision Index, marked “in review”. That ranking mixes Turion’s own public-suite run with the board’s official scores. We do not combine boards, so until the maintainers publish the private-test result we treat it as a vendor claim.

Author claims (UnverifiedClaim)

  • Decision Index 0.3 public index 66.16 on Turion’s own run (source)
  • Median 116 / mean 288 / p80 352 ms per request, one RTX PRO 6000, FP8 (source)
  • LoRA r64, 450 steps, about 7.3 hours on one B300 (source)

Scores (from scratch, 29 Oct 2026)

  • Capability 7.1 = types 10 × 0.30 + calibration 4 × 0.25 + latency 6 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.00 + 1.50 + 1.60 = 7.10. Calibration 4: temperatures without a published ECE. Latency 6: author figures from one setup. Accuracy 8 (was 5): board Full 70.47, 8 × 70.47 ÷ 59.22 = 9.5 → cap 8.
  • Maturity 5.7 = availability 8 × 0.30 + docs 7 × 0.25 + integrations 3 × 0.25 + support 4 × 0.20 = 5.70.
  • Adoption 1 = engagement 1 × 0.35 + ecosystem 1 × 0.35 + production 0 × 0.30 = 0.70 → 1.

Official board result (29 Oct 2026)

publicado (OK do Fabricio 29/10 BRT). Where this conflicts with “Board status” above, this section wins.

The Decision Index 0.4.1 added Turion Jev 31B on 28 Oct (10:19 BRT): Full 70.47, #4 of 243; public 66.16 (the same number Turion reported); board-measured ECE 0.019, median 144 ms, p95 815 ms per row. Jev 1.13, the reference of the accuracy rule, is 59.22 (#55). JevBench also has a row since v1.7.1 (25 Oct, draw v1.7): Capability Score 80.9, Intelligence 70.9, calibration 90.9, p50 0.124 s self-hosted, $0.096 per 1,000 decisions (2.97× Jev, outside the 2× cost cap).

Score update (approved 29 Oct). Capability 6.5 → 7.1 = types 10 × 0.30 + calibration 4 × 0.25 + latency 6 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.00 + 1.50 + 1.60 = 7.10. Only accuracy is re-derived: 8 × 70.47 ÷ 59.22 = 9.5 → cap 8. Calibration and latency stay as scored on 25 Oct, although the board now measures ECE 0.019; raising them was left out of this decision. Maturity 5.7 and adoption 1 unchanged (106 downloads, 2 likes on 29 Oct).