strands-labs/strands-decider

What it solves

Strands Decider addresses the gap between large language models (LLMs) and traditional classification models. While LLMs are flexible but slow and expensive, and traditional classifiers are fast but require extensive training and domain expertise, Strands Decider provides a general-purpose decision model that is fast, easy to deploy, and capable of scoring or picking from options without needing to generate arbitrary text.

How it works

The project uses a pretrained decoder LLM torso (Qwen3.5-2B-Base) and replaces its language-modeling head with a small "pointer head" (approximately one million parameters). Instead of a decoding loop that generates text, it performs a single forward pass to score options by comparing the hidden state at an answer position against the hidden state of the option's last token. It supports three question types via a masked softmax:

• Noul: Yes/No questions.
• Choice: Picking one from a set of N options.
• Score: Rating an item on an ordered rubric.

It is natively multimodal, allowing it to process images as part of the state using a vision tower.

Who it’s for

It is designed for developers building agentic AI workflows who need fast, reliable decision-making components for tasks like model routing, tool selection, argument verification, triage, and guardrails.

Highlights

• High Speed: Median latency of 115ms on an RTX 3090.
• Calibrated Reliability: Provides confidence scores that are well-calibrated, helping users decide when to defer to a human.
• Efficient Inference: Multiple questions about the same text can be processed efficiently since the state is only loaded once.
• Hardware Flexible: Runs on NVIDIA GPUs, Apple-silicon Macs (via MLX), and CPUs.
• Multimodal Support: Can process images to answer decision-based questions without retraining.