Purpose
Strands Decider is a small, fast decision model, officially described as a "system one" model, aimed at agentic AI workflows. It does not do open-ended text generation, but instead focuses on two types of tasks: making a choice among a given set of options, and scoring content on a scale. Officially, it is positioned to fill the gap between LLMs and traditional classification models: faster to respond than LLMs, yet unlike traditional classifiers it does not require time-consuming and specialized training. It is especially suited for use with agentic workflows built with the Strands Agents SDK.
Main capabilities
- Three types of questions: Choice (choosing among several named options, for example "What language is sihamba ngokushesha? English, Zulu, or Dutch"), Noul (yes/no binary probability, for example "Are there three r's in strawberry"), Score (scoring on an ordered scale, for example "How positive is this sentence? Between 0 and 1").
- Confidence: every decision comes with a calibrated confidence. The materials state that on unseen short classification tasks, answers with confidence 0.9 or above are about 95% correct; below that value, one should confirm or ask a human. Officially, frontier LLM inference APIs do not expose an equivalent capability.
- Composable questioning: multiple questions can be placed in a single command, loading the state only once, which is more efficient.
- Server mode: it can run as a server, with questions asked via HTTP requests.
- Multimodal (specific models): Qwen3.5-2B-Base is natively multimodal; when using --vision, the server retains the vision tower, and requests can carry base64 images in state. The released checkpoints (v1, and the earlier hobson-v21 and v19) can answer image questions without image training.
How to use
Officially, the way to get started is through the strands-decider CLI. On Apple silicon Macs, --device mlx can run the model via MLX, 1.4 to 1.6 times faster than MPS; this feature requires the mlx extra, which is officially said to be released with the next version, and before then it can be installed from a cloned repository with pip install -e ".[mlx]". The cuda, mps, and cpu extras correspond to other devices. The example output comes from strands-decider-2B-qwen3.5-v1-2610, recorded using --device cpu.
Model Family and Naming
The model naming format is strands-decider-<size>-<base>-v<N>-<YYMM>: <size> is the base model scale (e.g., 2B, E4B), <base> is the base model family (e.g., qwen3.5, gemma4), v<N> is the release sequence number on the same base family, and <YYMM> is the release month. Earlier models retain their original names: strands-decider-2B-hobson-v19 and strands-decider-2B-hobson-v21. The models are mainly hosted under the StrandsAgents organization on Hugging Face, and the five v1 models are also under the amazon organization, with the same files and names.
Use Cases
The official list covers decisions within agent workflows: model routing (selecting the appropriate LLM for a task), tool selection (deciding which tool the agent should call next), parameter checking (validating parameters before a tool call is executed), triage (routing requests to the responsible team or queue), guardrails (grounding checks, safety classification, policy classification), evaluation (scoring model outputs at low cost, determining whether an answer is sufficient), and hybrid agents (letting the LLM make difficult decisions and the decider make mechanical decisions to reduce cost and latency).
Limitations and Known Facts
- The materials state that the Gemma 4 models are text-only, and --vision is only used for Qwen3.5 models.
- Specific values for accuracy and calibration, the trade-offs among different model sizes (differences among 12B, 26B-A4B, E4B, and E2B in accuracy, calibration, memory, and latency), and the sources of training data are explained in the official documentation; the materials give a score of 0.836 for v1 on 1,996 held-out code tasks, and a median latency of 115 milliseconds for the 2B model on an RTX 3090.
- One training run takes about 11 hours on a single RTX 3090, or about 1 hour 10 minutes on eight H100s with FAST=1.
- The project uses the Apache License 2.0.
Not Yet Provided
Apart from the above, other features, fees, pricing, platform support, performance details, or usage steps are not yet provided in the materials.
