AWS Strands Labs Releases Open Source Decision Model

AWS Strands Labs has released Strands Decider 2B, an open-source decision model that focuses on making choices rather than generating text. This model processes a given state and typed questions to return a selected option, a yes/no probability with calibrated confidence, or a score on a rubric. The Strands Decider 2B model features 1.9 billion parameters and is designed for local execution on various hardware, including CPUs, consumer GPUs, and Apple silicon Macs. Its deployability is geared towards local and self-hosted environments, with model weights available on Hugging Face under the Apache-2.0 license. Users can install the model via pip, which provides both a command-line interface (CLI) and an HTTP server. The bundled HTTP server defaults to binding on 127.0.0.1 without authentication, necessitating the implementation of a custom authentication layer for production use. Currently, no hosted inference providers offer support for this model.

Decision models, also referred to as System One models, have emerged as a distinct category in the AI landscape following the recent launch of Jev by TypeSafe AI. Unlike large language models (LLMs) which can produce a wide range of arbitrary outputs, decision models are specifically engineered to select from a predefined set of options or to rate information on a scale. Strands Decider supports three primary question types: 'choice,' which selects one option from a list of N possibilities; 'noul,' which provides a binary probability between 0 and 1; and 'score,' which assigns a level based on an ordered rubric. Each response generated by the model is guaranteed to be from the allowed options and is accompanied by a confidence level. The development team acknowledges that Strands Decider is less capable than reasoning models when dealing with complex problems and is not suitable for tasks such as coding, general chat, or summarization.

The architecture of Strands Decider 2B is derived from the Qwen3.5-2B-Base model, with its language-modelling head removed and replaced by a smaller pointer head. This pointer head, comprising approximately 1 million parameters, compares the hidden state at the <answer> position with the hidden state of each option's last token. This design allows for a single forward pass to yield a result without requiring a decoding loop. The core of the model, referred to as the torso, utilizes a rank-16 LoRA (Low-Rank Adaptation) technique, while the pointer head operates in FP32 precision. The flexibility of the system allows label sets to be defined by the request, meaning there is no inherent limit to the number of options that can be presented. The released checkpoint is designated as v19. A significant advantage of this architecture is the efficiency of querying multiple questions against the same text; the state is processed only once, and each subsequent question incurs only the cost of its own tokens, making it cost-effective for iterative decision-making processes.