Quick Facts
- Amazon released Strands Decider 2B on October 1, 2026, under an Apache-2.0 license on Hugging Face and GitHub.
- The 1.9-billion-parameter model makes decisions in under 100ms on common hardware, including the NVIDIA RTX 3090.
- It ranks first on JevBench among public models that ship with a full training recipe, beating closed competitors including TypeSafe's Jev.
Amazon released a new open source model on October 1 that does not generate text. Strands Decider 2B, built by the Strands Labs team inside AWS, takes a list of options, scores them, and returns a selection with a calibrated confidence score in a single forward pass.
The model is built on Qwen3.5-2B-Base. AWS fine-tuned it with LoRA and added a roughly 1-million-parameter pointer head that scores each option's hidden state directly, skipping the decoding loop that slows standard language models. The result is a median latency of 106 milliseconds on an RTX 3090 and around 150 milliseconds on an M3 MacBook.
AWS Distinguished Engineer Marc Brooker led the project. He told TechCrunch the model is ideal for a specific problem in agent pipelines: "Given where I am right now, what should I do next?" That question, he argued, does not require a full LLM. It requires a fast, accurate classifier.
The model integrates into AWS's Strands agent framework through an intervention system. When an agent considers an action, Strands Decider 2B can approve it, deny it, request confirmation, or flag it for feedback. Generative calls still route through Amazon Bedrock, so the decision model acts as a cheaper gate before a full generative call fires.
VentureBeat reports the model hit 72.3% accuracy on the JevBench public set and a perfect score on the benchmark's easy tier. Among public models near the 2-billion-parameter mark, it ranks second overall. Among those that also publish a full training recipe, including data and scripts, it ranks first.
The key rival is TypeSafe AI's Jev model. Jev is a closed, hosted API with an undisclosed base model and no published training recipe. TypeSafe reports latency of 70 to 500 milliseconds depending on load. Strands Decider 2B is self-hosted, Apache-2.0 licensed, and ships with everything needed to reproduce the training run.
Brooker said rivals "might be underestimating the difficulty of making the models actually smart." He also said he does not expect frontier labs to dominate the category, because specialized niche models of this type cost hundreds or thousands of dollars to build, not millions.
One calibration result stands out for production use: on unseen short classification tasks, answers the model scored at 0.9 confidence or higher were correct about 95% of the time. AWS recommends confirming or escalating any decision that falls below that threshold.
Installation is straightforward. Developers run pip install strands-decider to get a CLI and an HTTP server. AWS warns that the bundled server binds to 127.0.0.1 with no authentication layer, so teams deploying to production must add their own.
The release points to a broader shift in how engineering teams may architect agentic systems. Small, fast decision models handle routing and gating. Large models handle generation. That split could reduce costs significantly in pipelines where most decisions are binary or classificatory, and only a fraction of calls need full language model output.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.