Strands Decider: Experimental Local Decisions and Explicit Vision Mode
Strands Decider is an experimental local decision-model project. Its current reference is Hobson v21; image handling needs the documented vision mode and dependencies.
Checked October 6, 2026 · Documentation review, no hands-on product test
A local project, not a managed cloud service
The current repository identifies StrandsAgents/strands-decider-2B-hobson-v21 as the reference checkpoint. It is an approximately 1.9B-parameter model using a Qwen3.5-2B-Base adapter. The repository and checkpoint declare Apache-2.0; they do not establish an included managed cloud endpoint. Current project README · Official v21 model card.
CLI and self-hosted /v1/systemone workflows expose Choice, Noul and Score. CUDA, MPS and CPU paths are documented; MLX is separately selected. Main-branch device extras are not all in the published PyPI 0.1.0 package. Match documentation to the version installed. Inference documentation.
Vision is an explicit operating mode
Image support requires --vision, the vision extra/Pillow and Transformers 5.18 or later. Documentation allows up to four base64 PNG/JPEG/WebP/GIF images and scales images down, never up, to a longest side of at most 448 pixels. A non-vision server rejects images with HTTP 422; its health response exposes the mode. No image-specific training is claimed. Vision mode and limits.
Check mode and preprocessing before evaluating screenshot decisions. A base model name alone does not tell you whether your running server accepts an image.
Version the evidence, not just the model name
The repository's 115 ms RTX 3090 JevBench figure is attached to v19. It is not a newly measured v21 latency result. The current recipe uses a 4,096-token window; validate your exact runtime configuration rather than treating an evaluation setting as a universal account allowance. Versioned research notes.
Suggested local brief: route queue items to a known handler, including a defer option. Keep missing inputs, out-of-domain cases and inconsistent answers in a fallback path. Validate caller permissions and tool arguments independently before running any selected action.
No model was installed or tested for this review. Compare operating responsibilities in the Decision Models selection guide.
Evidence and next step
Sources are linked beside the claims they support. Workflow suggestions are editorial examples, not completed tests; no account, installation, paid generation or benchmark was used for this profile.
Start with the official product or project and confirm the exact access, version and terms needed for your task. How we review.