Strands Decider Release - Held Out Code Accuracy: leaderboard
Strands Decider Release - Held Out Code Accuracy: Native task score. Official Strands decider release table: four explicitly linked gemma4-v1-2610 trained decision checkpoints, bf16 weights. JevBench correct counts out of 231, Brier error 0-2, held-out short/code accuracy fractions and native Decision Index 0-100. Latency and memory omitted. The separate Qwen3.5-2B checkpoint is not the Gemma E2B model.
Metric: Native task score. Source: raw.githubusercontent.com. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | strands-decider-12B-gemma4-v1-2610 | 0.87 | |
| 2 | strands-decider-26B-A4B-gemma4-v1-2610 | 0.86 | |
| 3 | strands-decider-E4B-gemma4-v1-2610 | 0.82 | |
| 4 | strands-decider-E2B-gemma4-v1-2610 | 0.77 |
Interactive version: theaggregate.ai/benchmark?slug=strands-decider-release-held-out-code-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.