AWS releases open-source Strands Decider 2B model

AWS open-sourced Strands Decider 2B, a 1.9B local model that picks options with confidence scores in about 115 ms.

Amazon Web Services’ Strands Labs released Strands Decider 2B (checkpoint v19), an Apache-2.0 decision model of about 1.9 billion parameters that selects among predefined options rather than generating free text. MarkTechPost says it is derived from Qwen3.5-2B-Base, replacing the language-model head with a roughly one-million-parameter pointer head and a rank-16 LoRA, and answers choice, yes/no, or rubric-score questions in one forward pass; TechCrunch instead names the base Qen3.5-2B and calls it inspired by TypeSafe’s Jev. On public JevBench v1 it scored 0.723 accuracy (167 of 231), with a Brier score of 0.342 and expected calibration error of 0.052, and median latency was 115 ms on an RTX 3090. Marc Brooker said AWS customers wanted a cheaper, lower-latency step for agent workflows that do not always need a frontier LLM, and that a homebrew version briefly led Jevbench for its size before release. TechCrunch also reports that TypeSafe’s Jev prompted many similar decision models. The bundled server is described as binding to localhost without authentication.

• AWS Strands Labs released Strands Decider 2B (checkpoint v19), an Apache-2.0 open-weights model of about 1.9 billion parameters.
• MarkTechPost says it is built from Qwen3.5-2B-Base with a roughly one-million-parameter pointer head and a rank-16 LoRA; TechCrunch names the base Qen3.5-2B.
• It returns choices, yes/no probabilities, or ordered rubric scores with confidence, not free text, in one forward pass.
• On public JevBench v1 it scored 0.723 accuracy (167 of 231), with a Brier score of 0.342 and expected calibration error of 0.052.
• Median latency was about 115 ms on an RTX 3090.
• Marc Brooker said customers wanted a cheaper, lower-latency agent step, and that a homebrew version briefly led Jevbench for its size before release.
• TechCrunch says the model is inspired by TypeSafe’s Jev, which prompted similar decision models.
• The bundled server is described as binding to localhost without authentication.