Strands released Strands Decider 2B, a small open source decision model. It picks between given options and assigns reliability scores. It trades text generation for speed and low latency. It cannot handle complex reasoning or coding tasks. It runs locally on consumer hardware in tens of milliseconds. The model is based on Qwen3.5-2B with a custom pointer head. It ranks 3rd among 2B-class models on JevBench accuracy and calibration. Latency scales linearly with task size. It is released with weights, training data and scripts on GitHub and Hugging Face. It suits model routing, tool selection, guardrails and hybrid agents. A CLI and Strands agent examples are provided for testing. The repo includes full training history and architecture details.
What changed
Decision models were not widely available as open source, small, local options for fast experimentation.
What it unlocks
Running fast, low-latency decision tasks locally without text generation or complex reasoning.
- 2B parameters
- 115ms median latency on RTX 3090
- 153ms median latency on M3 MacBook
- 3rd of 33 in 2B class on JevBench
What you need to act on it
- local CPU or GPU
- Python environment
Sources