Model explainer · independent

Strands Decider 2B, explained

Strands Decider 2B is a small open decision model released by AWS Strands Labs on 1 October 2026. It does not generate text. It scores a fixed set of options in one pass and returns a selection with a calibrated confidence value, which makes it a cheap gate inside an agent loop.

It is notable for two reasons: it runs on a local CPU or GPU, and AWS published the training data and scripts alongside the weights — the most reproducible artifact in this category so far.

How it is built

The team took a pretrained Qwen3.5-2B torso and removed its language-modeling head, which is the part that predicts the next token. In its place they added a small pointer head — just over a million parameters — that scores each offered option against the hidden state at the answer position. A rank-16 LoRA adapter fine-tunes the torso.

AWS describes the released model as v19 of that architecture and documents the earlier iterations, including a slot head that performed significantly worse. Removing generation is the whole point: the model cannot write prose, so it also cannot drift out of format.

Where it sits in an agent loop

AWS frames the model as an answer to "given where I am right now, what should I do next?" — a question that does not need a full LLM. In the Strands agent framework it acts as an intervention: when an agent considers an action, the decider can approve it, deny it, request confirmation, or flag it for feedback. Generative calls still route through a larger model.

  • →Model routing: decide whether a request needs a frontier model or a cheap one.
  • →Tool selection: pick the next tool from a fixed set.
  • →Guardrails: check an action against a policy before it runs.
  • →Memory and context management: keep or drop a piece of state.
  • →Evaluations: grade an output against a rubric with a calibrated score.

Reading the published numbers

AWS reports the model is competitive on accuracy and calibration for its size. Coverage of the launch cites 72.3% on the JevBench public set and a perfect score on the easy tier, with a third-place finish among 2B-class models; the same coverage notes it ranks first among public models that also ship a full training recipe. All of these are vendor or launch-coverage figures, not independent audits.

Latency figures vary by source: roughly 106ms median on an RTX 3090, around 115ms in AWS's own recap, and about 150ms on an M3 MacBook. Different hardware, different measurement setup. One calibration result is worth noting for production use: on unseen short classification tasks, answers the model scored at 0.9 confidence or above were correct about 95% of the time. AWS recommends confirming or escalating anything below that threshold.

  • →The bundled HTTP server is documented as binding to 127.0.0.1 with no authentication layer — add your own before exposing it.
  • →The single-pass design is explicitly described as worse at complex reasoning than a reasoning model.
  • →No published claim of Jev-API compatibility was found. The question shape is similar; verify the schema before reusing a client.

Frequently asked questions

Is Strands Decider a Jev clone?

+

It belongs to the same category and launch wave, and AWS explicitly credits TypeSafe Jev with starting it. But it is an independent architecture with a published training recipe, not a copy of Jev weights — Jev weights have never been released.

Can it run on a laptop?

+

AWS says it is sized for a local CPU or GPU. Launch coverage reports roughly 150ms median on an M3 MacBook and around 106–115ms on an RTX 3090. These are published figures from different setups, so measure your own workload.

Does Strands Decider generate text?

+

No. The language-modeling head was removed. It scores offered options and returns a selection with a confidence value. AWS lists summarization, chatbots, and coding as tasks the approach is unsuited for.

Is it Jev-API compatible?

+

No compatibility claim was found in the materials reviewed here. The two use a similar state-plus-questions idea, but the request and response schemas are published separately. Do not assume an existing Jev client works without testing.

Where do the weights and training recipe live?

+

AWS says the code is on GitHub and the weights on Hugging Face under Apache 2.0, including the training data and scripts used to build the model. Check the Strands Labs repositories for the current paths.

Keep reading

Sources

Reviewed 2026-11-13. Vendor- and project-reported figures are labelled as such; none were independently reproduced by this site.

Model explainer · reviewed 2026-11-13 · first published 2026-11-13