AWS Strands Decider 2B: Open-Source Decision Model
AWS Strands Labs released Strands Decider 2B on 1 October 2026, describing it as a small open-source decision model optimized for fast experimentation and local development that picks among a set of options and returns calibrated confidence.
Tool Overview
Features, steps and FAQ below
Features
- ✓ AWS says it is a ~2 billion parameter model suitable for running on a local CPU or GPU that can return answers to meaningful questions in tens of milliseconds
- ✓ Architecturally it takes a pre-trained Qwen3.5-2B torso, removes the LM head and replaces it with a pointer head of just over a million parameters, with the torso fine-tuned via a rank-16 LoRA adapter
- ✓ Decision models do not generate text; they pick among a set of options or assign a 0-1 score, and give each decision a high-quality reliability score
- ✓ AWS says it performs well on accuracy and calibration on JevBench's public set, ranking 3rd of 33 in the 2B class and 1st of 30 excluding the just-over-2B models
- ✓ On latency, AWS says it makes local decisions in a median of around 115ms on an Nvidia RTX 3090 and around 153ms for small tasks on an M3 MacBook
How to Use
- Install the CLI with pip install strands-decider
- Use strands-decider ask with --state and --choice to let the model choose among the options you provide
- Get the code from GitHub (strands-labs/strands-decider) and weights from Hugging Face (StrandsAgents), with all training data and scripts included
- Integrate it into a Strands agent via the InterventionHandler before_tool_call hook for pre-tool-call validation or routing
FAQ
What is Strands Decider 2B?
A small open-source decision model released by AWS Strands Labs on 1 October 2026, which AWS says is optimized for fast experimentation and local development, picking among options or scoring instead of generating text.
Is it open source?
AWS says it released strands-decider-2b as open source on GitHub (strands-labs/strands-decider), with weights on Hugging Face (StrandsAgents) and all training data and scripts included.
How fast does it run?
AWS says it runs on a local CPU or GPU, with a median latency of around 115ms on an RTX 3090 and around 153ms for small tasks on an M3 MacBook.
How is it different from an LLM?
AWS says decision models trade flexibility for speed and stronger capability at a given size, always answer from the chosen options, are unsuited to coding, chatbots or summarization, and excel at routing, tool selection, evaluations, guardrails and policy classification.
How do I get started?
AWS says the easiest way is the strands-decider CLI: pip install strands-decider, then use strands-decider ask with a state and choices.