Amazon Web Services on October 1, 2026 released Strands Decider 2B, a small open-source decision model built to choose among predefined options and attach confidence scores for agent workflows, TechCrunch reported. The model ships from Strands Labs, AWS’s experimental agent toolkit group, and is available now for local use.
Decision models sit in a different lane from chat LLMs. Instead of generating free-form text, they score a closed set of answers – for example which team should handle a ticket, whether a tool call is premature, or a yes/no guardrail check – and return calibrated confidence, the Strands Agents blog said. In exchange for that narrower output space, the models are designed for lower latency and more reliable structured steps inside agent pipelines.
Amazon distinguished engineer Marc Brooker told TechCrunch the project started after he saw TypeSafe’s Jev and built a homebrew take that briefly topped the Jevbench ranking for models of its size. AWS customers’ agentic workflows, he said, often need a workflow step that answers “what is the next thing for me to do here” without paying for a full frontier LLM call every time. Strands Decider fills that role with closed-domain answers and confidence scores at lower latency and potentially lower cost, Brooker said.
Under the hood, Strands Decider keeps the torso of Alibaba’s Qwen3.5-2B, removes the language-model head that predicts next tokens, and adds a small pointer head that scores supplied options, Strands said. The release includes code on GitHub, weights on Hugging Face, and the training data and scripts used to build the model. Strands reports median local decision latency around 115 milliseconds on an Nvidia RTX 3090 and about 153 milliseconds for small tasks on an M3 MacBook, with answers on short tasks in tens of milliseconds. On JevBench’s public set, Strands says the 2B checkpoint ranks third of 33 models in its size class.
Early uses Strands flags include model routing, tool selection, evaluations, guardrails, memory and context management, and policy classification, including hybrid agents that keep hard decisions on an LLM and hand rote checks to the decider. TypeSafe CEO Diogo Almeida told TechCrunch his team is focused on making future models smarter and does not yet see real competition from the current batch of clones. VentureBeat separately covered the same AWS release as an open, self-hostable decision layer rather than a hosted API.
This brief covers the Strands Decider 2B model release for agent decision steps. It is not Amazon’s reported Nvidia Grace Blackwell SPV leaseback story.