Amazon Releases Open Strands Decider 2B for AI Agent Decisions

Amazon’s Strands Labs has released Strands Decider 2B, a roughly 2-billion-parameter model fine-tuned from Alibaba’s Qwen base to answer bounded questions with choices and probabilities in one pass. [1][2] The model is free to download from Hugging Face under the permissive Apache 2.0 license, although users must provide the hardware or cloud resources needed to run it. [2] Amazon says developers can download its training materials and use the model to route requests, select tools, evaluate outputs, or review an agent’s actions. [2]

AWS says the model starts from Qwen3.5-2B, replaces the next-word prediction component with a small pointer component, and scores supplied answer options directly. [2] Developers used a rank-16 LoRA update and trained the new scoring component, which AWS says adds just over one million parameters. [2] Strands Decider 2B is part of Strands Labs, AWS’s experimental agent-development project, and is positioned as a competitor to TypeSafe AI’s Jev. [1][2]

By separating bounded choices from prose generation, Strands Decider 2B gives agent builders a specialized checkpoint for deciding whether actions should proceed, while generative models can remain focused on writing, coding, and complex reasoning. Compared with TypeSafe AI’s Jev, Amazon’s entry pairs downloadable weights and training materials with an Apache 2.0 license, making infrastructure and workflow fit important evaluation criteria. What the sources don't address: how Strands Decider 2B’s accuracy, latency, and operating cost compare with Jev in independent production tests.