Strands Decider 2B: a small, open-source, decision model

Strands Decider 2B is a 2-billion-parameter, open-source decision model built to make fast, reliable choices among a fixed set of options rather than generate free-form text. It belongs to the emerging class of “system one” models that trade generative flexibility for speed, deterministic selection, and calibrated confidence scores - advantages that suit agentic workflows, parallel querying, and low-latency local development. Because it cannot generate arbitrary text, it is not intended for tasks like long-form chat, coding, or summarization, but it excels at yes/no, multi-choice, and scalar scoring tasks where a reliability estimate and quick turnaround matter.

Architecturally, the model reuses a Qwen3.5-2B torso with the language-model head removed and replaced by a small pointer head (~1M parameters) that scores candidate answers; the torso was fine-tuned with a rank-16 LoRA adapter. This is the v19 iteration after earlier experiments with a slot head that underperformed. On JevBench’s public set it ranks highly for accuracy and calibration (3rd of 33 in the 2B class, 1st excluding slightly larger models) using accuracy and Brier score metrics. Latency is low: median local decision time ~115 ms on an RTX 3090 and ~153 ms on an M3 MacBook for small tasks, scaling roughly linearly with task size. The full code, weights on Hugging Face, training data, and scripts are available for reuse and further experimentation.