Download
Strands Decider is a compact System One decision model built for fast classification and scoring inside AI agent workflows. Unlike a generative LLM, it chooses among options or rates inputs on an ordered scale without decoding free-form text. Its first model has 1.9 billion parameters and replaces the language-modeling head of Qwen3.5-2B-Base with a small pointer-based decision head. It supports yes-or-no, multiple-choice, and score questions while returning calibrated confidence values. Shared-prefix caching makes multiple questions about the same state inexpensive. The package includes a CLI, HTTP server, training pipeline, evaluation tools, and a worked Strands Agents SDK integration. Inference runs on CUDA, CPU, Apple MPS, and MLX.
Features
- Yes-or-no, choice, and score decisions
- Calibrated confidence values
- Single-pass non-generative inference
- Shared-prefix caching for repeated questions
- CUDA, CPU, MPS, and MLX execution
- CLI, HTTP API, training, and Strands agent integration
Project Samples
