Strands Decider 2B is a roughly 2B-parameter model on a Qwen3.5-2B base that VentureBeat reports runs at a 106 ms median on an RTX 3090. It suits teams that gate or route agent steps locally and want an open alternative to Jev.
What changed
Strands Decider 2B is an open decision model from Amazon's Strands Labs, released 1 October 2026 under Apache 2.0. It scores developer-supplied answer options with a confidence instead of writing text, and Strands Labs reports 72.3% accuracy on the JevBench public set (v19 checkpoint).
Provider: Amazon · Family: Strands Decider
Input modalities: text, image · Output: text
About Strands Decider 2B
Strands Decider 2B is a decision model released on 1 October 2026 by Strands Labs, the open-source group at Amazon behind the Strands Agents SDK. A decision model is not a chat model. It takes a state and one or more typed questions, and returns a probability for each allowed answer instead of writing text. Strands Labs describes the three question types as a yes/no question, a choice among N options, and a level on an ordered scale. The class is also called a "system one" model, and the closest commercial product is Jev from TypeSafe AI.
The architecture is a plain edit of a small language model. According to the Hugging Face card, the 2B checkpoint is a LoRA adapter on Qwen3.5-2B-Base plus a small readout head that scores the options of each question. Press coverage puts the adapter at rank 16 and the head at about one million parameters, and says the next-word prediction head was removed. The result is one forward pass per decision with no generation step. The released reference checkpoint is named strands-decider-2B-hobson-v19, and the repository has since added strands-decider-2B-qwen3.5-v1-2610 on the same base.
Scores come from Strands Labs and should be read as vendor-reported. The v19 card lists 72.3% accuracy on the JevBench public set (167 correct on its 231 tasks), a Brier score of 0.348 and an expected calibration error of 0.050. The README for the newer qwen3.5-v1-2610 checkpoint lists 180 correct on the same 231 tasks with a Brier score of 0.280. JevBench is versioned and scores across versions are not comparable, so every figure here names the version it came from. VentureBeat, quoted by AI Weekly, reports that a competing open model, Mapika decider-2b v11, reaches 76% on the benchmark. Internal held-out sets on the v19 card range from 64.1% on short tasks to 88.4% on MuSiQue.
Every answer carries a confidence value, and the README says that on short classification tasks the model has not seen, answers at 0.9 confidence or higher are right about 95% of the time. The same README tells readers to measure on their own traffic before trusting a threshold. The card lists the weak spots plainly: a changed question often returns the same answer when the state and options are fixed, long multi-step documents are the hardest tier, and the score and yes/no question types transfer poorly to rubrics unlike the training mix.
Speed is the reason to use it. VentureBeat reports a 106 ms median and 296 ms p95 on an NVIDIA RTX 3090 and roughly 150 ms median on an M3 MacBook for small tasks. These are secondary reports of Amazon's figures, not independent runs. The model runs on CPU, CUDA GPUs or Apple silicon (the README says the MLX path is 1.4 to 1.6 times faster than MPS). Qwen3.5-2B-Base is natively multimodal, and the card's repository notes say image states work when the server starts with the vision flag, with no image training in the published checkpoints.
There is no per-token price because there is no hosted API from Amazon. The weights, training data inventory and scripts are released under Apache 2.0, so the cost is your own hardware. The recipe used one p5.48xlarge host with H100 GPUs for about 4,204 seconds of wall clock, and the card says a retrain takes about 11 hours on one RTX 3090. For a hosted alternative with a per-token price, see Cloudflare Clef, which sells decision models on Workers AI.
Deployment is local. Install the strands-decider package, run its ask command for one-off questions, or run its serve command to expose an HTTP endpoint on 127.0.0.1. The card warns the server has no authentication and is meant for local experiments, so put your own gateway in front of it. The code ships examples for the Strands Agents SDK, and the weights load with Hugging Face PEFT. Local runners such as Ollama and LM Studio are built for text generation and are not the supported path here. Our guide on how to run AI locally covers the general hardware questions.
Who should use it: teams that route tickets, pick tools, gate agent steps or grade outputs with an LLM prompt today and want a typed answer with a confidence in about a tenth of a second on hardware they control. Who should not: anyone who needs open-ended reasoning or text, since the model cannot write. It also trails larger decision models on Strands Labs' own table, where the Gemma 4 based 12B and 26B-A4B variants score 200 and 207 on the same 231 tasks. Teams that want a managed service should compare Jev and Cloudflare Clef, and teams choosing a small general model can look at Gemma 4 12B. Early coverage also flagged that adversarial text in an agent's input can influence a verdict, and that Amazon has not shown that self-hosting beats a hosted price once hardware and operations are counted.
Training data is listed source by source in the repository. The card names ContractNLI and MuSiQue from their authors, public Hub datasets, synthetic rows that open-weight models generated and checked, and the output distributions of a frozen Qwen3.5-4B teacher. The HokAI entry covers the 2B family; the larger Gemma 4 based checkpoints in the same repository are separate models. Amazon engineer Marc Brooker is quoted in coverage as the person who started the project after seeing the class of models.
Pricing
Amazon runs no hosted API for it. Running it costs only your own CPU or GPU time.
Key Features
- Typed decision outputs: Three question types (yes/no, choice, score) each return a probability, with a confidence on choice and score answers.
- Pointer head, no generation: A rank-16 LoRA and a small readout head replace the next-word head on a 2B base model.
- Calibrated confidence: Choice and score answers carry a confidence value that you can threshold before acting.
- Local serving: Run it on CPU, CUDA or Apple silicon, or expose it on 127.0.0.1 with the serve command.
- Reproducible recipe: Weights, a data inventory and training scripts ship together, so a team can retrain the model.
Pros
- Open weights, a data inventory and scripts, all published.
- One forward pass per decision keeps latency around a tenth of a second on a consumer GPU (vendor-reported).
- Every answer carries a confidence you can threshold.
Cons
- Cannot generate or explain text.
- Weak on long multi-step documents per the model card.
- Scores are self-reported and trail the larger Gemma 4 based variants in the same repository.
Frequently Asked Questions
What does Strands Decider 2B cost to run?
Amazon sells no hosted API, so there is no per-token price. The weights are free under Apache 2.0 and you pay for your own CPU or GPU. For comparison, Jev lists $0.042 per 1M input tokens with no output fee, according to AI Weekly.
How does Strands Decider 2B score against other decision models?
The v19 checkpoint reports 167 correct on JevBench public tasks and the newer qwen3.5-v1-2610 README lists 180. VentureBeat says Mapika decider-2b v11 reaches 76%, so the newer checkpoint is the fairer comparison. Always name the JevBench version, since versions are not comparable.
Is Strands Decider 2B really open source?
Yes. Weights, a data inventory and training scripts are published under Apache 2.0 on Hugging Face and GitHub. The base is Qwen3.5-2B-Base, and the data list includes sources with their own licenses recorded in data/sources.md.
Does Amazon see what I send to Strands Decider 2B?
Not unless you build a service yourself. The model runs locally through the strands-decider package, so inputs stay on your machine. The built-in server has no authentication, so keep it on localhost.
Should I pick Strands Decider 2B or a hosted decision model?
Pick it when you need weights you control, offline use or a reproducible recipe. Choose a hosted option such as Jev or Cloudflare Clef when you want a managed endpoint and no GPU to run. Neither writes text, so keep an LLM for generation.