Amazon Web Services has released Strands Decider 2B, a 2-billion-parameter "decision model" that sorts between predefined options and returns a confidence score instead of generating text — the company's answer to TypeSafe AI's Jev, the model that kicked off the category earlier this month.
The model was published by Strands Labs, the AWS group building tools and protocols for deploying AI agents. It is fully open source: code on GitHub, weights on Hugging Face, and — unusually — the training data and scripts used to build it. AWS says the 2B model is small enough to run locally on a CPU or GPU and returns answers in tens of milliseconds.
WHAT A DECISION MODEL IS
Unlike a large language model, which can produce arbitrary output, a decision model is designed to pick from a supplied set of options ("Is the phrase 'sihamba ngokushesha' English, Zulu or Dutch?") and attach a numerical reliability score to that choice. In exchange for this loss of flexibility, AWS says, such models are faster and more capable at a given size, always return an answer drawn from the offered options, and run with very low latency. The trade-off is that they are worse at complex reasoning and cannot generate text, making them unsuited to coding, chatbots or summarization.
ARCHITECTURE AND RESULTS
Strands Decider 2B is built on the "torso" of Qwen3.5-2B: the language-model head is removed and replaced with a small pointer head (just over one million parameters) that scores each offered option against the answer position. The torso is fine-tuned with a rank-16 LoRA adapter. It is the nineteenth iteration of the architecture, and the second major one — the first used a slot head that performed significantly worse.
On JevBench's public set, AWS reports the model ranks third of 33 in the 2B class on accuracy and calibration, and first of 30 when models just over 2B are excluded. Median latency is about 115ms on an Nvidia RTX 3090 and roughly 153ms for small tasks on an M3 MacBook.
ORIGIN AND CONTEXT
AWS distinguished engineer Marc Brooker built the first version after seeing Jev, and it briefly took the top spot on the Jevbench ranking for its size. AWS engineers then cleaned it up for release under Strands Labs. Brooker told TechCrunch the need surfaced in conversations with customers whose agentic workflows did not always justify the cost or latency of a full LLM: a decider makes "a perfect" step for a workflow asking "what is the next thing for me to do here?"
The release lands the same week OpenAI announced a similar offering, its Decisions API, and after researchers produced dozens of comparable models since TypeSafe debuted Jev. TypeSafe CEO Diogo Almeida said he saw no serious competition yet: "The current batch seems more like ML people wanting to implement a cool architecture than a team deeply dedicated to making intelligence useful."
WHY IT MATTERS
AWS positions the class for model routing, tool selection, evaluations, guardrails, memory and context management, and policy classification — and for hybrid agents that use LLMs for the hardest decisions and deciders for routine ones, cutting cost and latency. The bet is that not every step of an agent's workflow needs a frontier model; a cheap, fast, locally runnable model that is reliably certain about small questions may be enough for many of them.
