Strands Labs Releases Open Decision Model Strands Decider 2B

Strands Labs released Strands Decider 2B, a small open model that picks among options in about 115 ms rather than generating text.

Strands Labs — identified as AWS Strands Labs in one report — released Strands Decider 2B, an open-weights model that selects among supplied options, answers yes/no questions, or returns rubric scores instead of generating free text. One source describes checkpoint v19 as a 1.9-billion-parameter model derived from Qwen3.5-2B-Base; the other calls it a 2-billion-parameter model built from a Qwen3.5-2B torso. Both agree the language-model head was replaced by a pointer head of about one million parameters plus a rank-16 LoRA, answering in a single forward pass. On public JevBench v1, one report gave 0.723 accuracy (167 of 231), a Brier score of 0.342, and expected calibration error of 0.052; the later report said it ranked third of 33 models in the 2B class on accuracy and Brier-score calibration. Median latency was 115 ms on an Nvidia RTX 3090, with one report also citing 153 ms on an M3 MacBook. Weights are Apache-2.0, with code, training data, and scripts on Hugging Face and GitHub; one source says the bundled server binds to localhost without authentication.

• Strands Labs (called AWS Strands Labs in one report) released Strands Decider 2B, identified as checkpoint v19.
• Sources disagree on size and base: 1.9 billion parameters from Qwen3.5-2B-Base versus a 2-billion-parameter model from a Qwen3.5-2B torso.
• The language-model head is replaced by a pointer head of about one million parameters and a rank-16 LoRA; it returns choices, yes/no probabilities, or ordered scores, not free text.
• On public JevBench v1: 0.723 accuracy (167 of 231), Brier score 0.342, and expected calibration error 0.052; also ranked third of 33 models in the 2B class for accuracy and Brier-score calibration.
• Median latency was 115 ms on an Nvidia RTX 3090; one report also cites 153 ms on an M3 MacBook.
• Weights are Apache-2.0; training data, code, and scripts are on Hugging Face and GitHub.
• One report says the bundled server binds to localhost without authentication.