AWS Strands Decider 2B: Fast Decision Model Hits 115ms

AWS Strands Labs put out a new open source model on October 7th that doesn't write a single word of text. Strands Decider 2B takes in a state, a typed question, and a set of allowed options, then hands back a choice, a probability, or a score, each with a calibrated confidence attached. At 1.9 billion parameters, it's light enough to run on a laptop CPU, a consumer GPU, or an Apple silicon Mac, and the team claims decisions land in about 115 milliseconds with an accuracy score around 0.7 on their benchmark.

The model slots into a category that's barely a month old. TypeSafe AI's Jev kicked off the "decision model" idea, sometimes called System One models, as a counterpoint to the generate-anything approach of large language models. Instead of open-ended output, these models restrict themselves to picking between fixed options or rating something on a rubric, trading flexibility for speed and predictability.

Strands Decider is available now through Hugging Face under an Apache-2.0 license, installable via pip with a command-line tool and a bundled HTTP server. The team is upfront about where it falls short, and the mechanics behind how it actually reaches a decision start with a fairly blunt architectural choice.