AWS Strands Labs has launched Strands Decider 2B, an open-source decision model designed to select from predefined options or rate inputs on a scale. The model, released as checkpoint v19, is available on Hugging Face under the Apache-2.0 license. Users can deploy it locally or via self-hosted environments using a CLI or HTTP server.
The model is built upon the Qwen3.5-2B-Base architecture, replacing the language-modeling head with a pointer head containing approximately 1 million parameters. It utilizes a rank-16 LoRA for the torso and runs the head in fp32. This architecture allows for a single forward pass without a decoding loop, achieving response times of approximately 115 ms.
In performance evaluations on the JevBench v1.4.2 board as of September 25, Strands Decider 2B ranked 3rd out of 33 models in the 2B class. The developers highlight that the model is not intended for complex reasoning, coding, or summarization tasks, but rather for specific functions like model routing, tool selection, and guardrails.
The model demonstrates high calibration, with answers at a confidence level of 0.9 or higher proving correct approximately 95% of the time on short classification tasks. The team suggests that users should manually confirm or escalate any decisions made with a confidence score below this threshold.
