Is Raw Qwen3 4B Instruct 2507 direct logits SOTA in Decision models?#389 in Decision modelsRanked #389 in Decision models.
How it scores here
Raw Qwen3 4B Instruct 2507 direct logits in Decision models
Primary benchmarks
- JevBench: 34.6 (CI 34.1–35.1) on Oct 8, 2026 (#116 of 134).
On the category's weighted score it is 86.4 points behind #1 karpo-decision-31B and 0.3 behind the next model up, Bekko System One v0 400M.
Nearby on the board
| # | Model | Category score |
|---|---|---|
| 386 | Raw Phi-4 mini direct logits | 14.4 |
| 387 | Crivo-150M v2 | 14.2 |
| 388 | Bekko System One v0 400M | 13.9 |
| 389 | Raw Qwen3 4B Instruct 2507 direct logits | 13.6 |
| 390 | Mola-150M | 13.6 |
| 391 | vLLM-SR d3-edge (0.6B) | 13.3 |
| 392 | Mola-68M | 13.0 |
Standings
Decision models
| # | Model | State |
|---|---|---|
| 1 | karpo-decision-31B | Co-SOTA |
| 1 | Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout) | Co-SOTA |
| 1 | GLiDE 27B | Co-SOTA |
| 1 | kudhab (Gemma 4 31B) | Co-SOTA |
| 2 | Quyet-1.0-Large (Chinh Nguyen, Gemma-4-31B decoder, option-letter logits) | Near-SOTA |
| 3 | REAX 1.3 | Near-SOTA |
| 4 | Logits Sigma-1 Duo | Near-SOTA |
| 5 | KnowLine 31B Gen2 | Near-SOTA |
| 6 | Strands Decider v2 31B | Near-SOTA |
| 7 | Copernicus-31B | Near-SOTA |
| 8 | decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted) | Near-SOTA |
| 9 | Copernicus-57B | Near-SOTA |