AWS Strands Labs releases Strands Decider 2B, an open decision model
AWS’s Strands Labs announced Strands Decider 2B on 1 October 2026. It is an open decision model built for agents made with the Strands Agents SDK: Qwen3.5-2B with the language head swapped for a small pointer head (about a million parameters) and a rank-16 LoRA on the torso. It answers Noul, Choice and Score questions with a probability per option.
Weights are on Hugging Face, and code, training data, scripts and a long research log are in strands-labs/strands-decider (Apache-2.0, 171 stars in three days). pip install strands-decider gives a CLI and a server that speaks /v1/systemone. An Ollama PR (#18748) is open.
The numbers (UnverifiedClaim). On JevBench’s public set, with JevBench’s harness, v19 scores 0.723 (167 of 231), Brier 0.342, ECE 0.052; AWS says that would place it 50th of 90 and first among models of 2B or fewer, but it is not on the board. Median latency is 115 ms (p95 299 ms) on an RTX 3090. The repo itself shows score answers are weak on unseen tasks (0.499).
How we list it. Catalog, 6.3 maturity / 7.7 capability / 14 adoption.