ENPT

Strands ships five v1-2610 decider checkpoints on Gemma 4 and Qwen3.5; the 26B-A4B claims 89.6% on JevBench public

Last updated:

Source · Model card: Strands Decider 2B

Strands (an AWS lab project) released five decision-model checkpoints on 9 October 2026, the first on Gemma 4: E2B, E4B, 12B and 26B-A4B, plus a 2B on Qwen3.5. Each is a LoRA adapter plus readout head, Apache-2.0, served by strands-decider serve on /v1/systemone.

  • Headline: 26B-A4B 207/231 (89.6%), Brier 0.162, on JevBench’s public split at a 4,096 window; 12B 200/231 (86.6%). Previous best, hobson-v21: 176/231 (76.2%). UnverifiedClaim.
  • Caveats: public items only, self-run; not on Benchmark Heaven’s board or Decision Index 0.3. The chart’s “beat earlier models on calibration” holds for Brier, not ECE (the 2B’s 0.072 is worse than v21’s 0.064). Gemma variants train on v21’s data; only the 2B drops its 20,000 commercially restricted rows.

Full table on the Strands decider page. Capability on that page goes from 7.7 to 8.1: the public-split result is counted as evidence (still UnverifiedClaim, not on any board).