AWS debuts Strands Decider 2B, a first lightweight decision model for accelerate agentic workflows

Amazon Web Services is pushing into a very specific corner of AI: models that don’t write, chat or explain, but simply choose. Its Strands Labs team has released Strands Decider 2B, an open-source decision model built to make rapid calls by avoiding text generation altogether. That matters because text is expensive in token use and slow in response time, and AWS is betting there’s a real place for a dedicated decision engine inside agentic systems.

Decision models, or “System 1 models,” are a different beast from the large language models that dominate the current AI conversation. They take a set of predefined choices, pick one, and attach a confidence score. No prose. No code. No image output. The upside is speed and low latency; the downside is obvious, too, because the model can’t explain itself in the way a normal LLM can.

AWS says Strands Decider 2B is its first serious shot at improving the category. The model sits on top of Qwen3.5-2B, but the usual LLM head has been replaced with a tiny custom pointer head containing about 1 million parameters. AWS also fine-tuned the torso with a rank-16 LoRA adapter so it can score hidden states of available choices against answer positions. The company says it has already iterated several times on the model, and today’s release is v.20.

The scale is deliberate. AWS picked 2 billion parameters because it thinks that size can run locally with less than 150 milliseconds of latency while still handling complex decisions. The company also says the model performs well on JevBench for accuracy and calibration compared with other open-source 2B models. The download is live on Hugging Face, with code, training scripts and examples on GitHub.

And AWS isn’t just aiming at one narrow demo. It wants developers to use the model for routing, tool selection, context management, guardrail enforcement and policy classification, then pair it with larger LLMs for harder reasoning. That hybrid approach is the interesting part: let the small model do the boring choices, and stop making the big model pretend every task needs a novel.