KnowLine-31B-Gen1 takes #1 on the Decision Index; Turion Jev 31B enters at #4
Source · Model card: KnowLine-35B-A3B Gen1
publicado (OK do Fabricio 29/10 BRT).
The Decision Index 0.4.1 added five commits of new rows on 28 Oct 2026, between 08:52 and 16:33 BRT (243 rows in the data of 28 Oct, 19:32 UTC). The first batch put KnowLine-31B-Gen1 from PelaAI at the top of the board.
| Decision Index 0.4.1 (data 28 Oct) | Full | Public | Rank |
|---|---|---|---|
| KnowLine 31B Gen1 (new) | 71.86 | 70.13 | #1 |
| Fastino GLiDE 1.2 | 71.52 | 71.99 | #2 |
| Sieve 31B v3 | 71.10 | 71.68 | #3 |
| Turion Jev 31B (new) | 70.47 | 66.16 | #4 |
| vLLM-SR d3-ultra | 70.18 | 71.57 | #5 |
| H2O Lightning 31B v1.1 (new) | 68.64 | 67.75 | #7 |
| OneShot 27B (new) | 65.74 | 65.41 | #28 |
| vera-large (new) | 65.26 | 67.75 | #29 |
| Jev 1.13 | 59.22 | 57.96 | #55 |
Ranks are by Full score. The lead is small: 0.34 points over GLiDE 1.2 and 0.76 over Sieve 31B v3, which entered at #2 a day earlier.
KnowLine-31B-Gen1 is an Apache-2.0 LoRA fine-tune of Gemma 4 31B (merged weights), published on 23 Oct. Its Full score is higher than its public score, the opposite of the 35B-A3B sibling this site already covers (Full 62.92, public 69.49 on the 0.3.2 board). The board measured ECE 0.032, a median of 204 ms and p95 1,822 ms per row. PelaAI says an AI agent ran the training loop end to end. The card’s other numbers (JevBench 90.5, new-domain tasks 74.9, zero-shot boxes) are the vendor’s own runs (UnverifiedClaim).
Turion Jev 31B, which this site scored on 25 Oct from the author’s public-suite run (66.16), now has an official line: Full 70.47, ECE 0.019, median 144 ms, p95 815 ms.
Score. The KnowLine page keeps its scores; its accuracy was already at the cap of 8. The Turion Jev 31B page moves from capability 6.5 to 7.1 under the accuracy rule (proposal).