Introducing Strands Decider 2B: a small, open source, decision model
Short summary
Strands released Strands Decider 2B, a 2-billion-parameter model designed to quickly pick between predefined options rather than generate arbitrary text. It runs locally in ~115ms and ranks third among 2B-class decision models on accuracy and calibration benchmarks.
- Decision models sacrifice flexibility for speed and reliability: they choose from given options and assign confidence scores, making them 100x faster than LLMs for simple classification tasks but unsuitable for text generation, coding, or reasoning.
- The 2B parameter size hits a sweet spot for local experimentation—small enough to run on consumer hardware, large enough to handle meaningful work like tool selection, routing, and guardrails in agentic workflows.
- Strands open-sourced the full model, training data, and 19 iterations of architecture refinements on GitHub and Hugging Face, explicitly inviting community contributions to the decision model space.
Full summary
Strands released Decider 2B, a 2 billion parameter decision model built for fast, local inference. Unlike large language models that generate arbitrary text, decision models pick between predefined options and assign confidence scores—think yes/no questions or choosing from a list. The tradeoff is clear: they're faster, more reliable at their specific job, and always stay within their option set, but they can't handle complex reasoning or generate text. Strands built Decider 2B by taking Qwen 3.5-2B, stripping out the text generation head, and replacing it with a pointer head that scores options against a hidden state. The model runs in about 115 milliseconds on an RTX 3090 and 153 milliseconds on an M3 MacBook, with latency scaling roughly linearly with task size.
Performance-wise, Decider 2B ranks third among 2B-parameter decision models on JevBench's public benchmark for both accuracy and calibration (using Brier score). It nails 100% of easy tasks, which matters because agents often need to make simple routing decisions repeatedly. The 2B size was deliberate—large enough to do real work but small enough to run on consumer hardware, lowering the barrier to experimentation. Strands released the full package open source: model weights on Hugging Face, training data, training scripts, and the complete version history showing nineteen iterations of refinement.
Early use cases show developers deploying decision models for model routing, tool selection, guardrails, and memory management. The hybrid pattern gaining traction pairs decision models with LLMs—let the small model handle cheap, fast yes/no calls and context checking, reserve the expensive LLM for genuinely hard decisions. Strands included an example agent that uses Decider 2B to catch bad tool calls before they execute, asking whether the proposed arguments actually came from the user or whether it's too early to make that call. The intervention system lets developers hook decision points into agent workflows with a few lines of Python.