Open source, self-hosted. There is no hosted API. You download the weights and run them on your own hardware. How open-source models are listed.
What Strands-Decider is
Strands-Decider 2B is an open-weights decision model from Strands Labs, which the Strands Agents documentation calls its experimental arm. It is aimed at the small decisions inside an agent: picking a model, choosing a tool, checking a tool call’s arguments, triaging a request, or screening for safety.
On the AWS link: no page says outright that AWS builds it. The evidence is that the repo’s security policy points to AWS’s vulnerability program, the PyPI package is owned by the strands-agents account, Amazon’s verified Hugging Face organisation mirrors the v1 models, and Microsoft’s benchmark labels the row AWS. This page calls the maker Strands Labs.
How it works
It takes Qwen3.5-2B-Base, removes the language-generation head, and adds a rank-16 LoRA adapter plus a pointer head of about a million parameters. That gives 1.9 billion parameters. The head scores each option in one forward pass, so nothing is generated. Questions are Noul (yes/no), Choice (one of N) and Score (a level on an ordered scale), served over a /v1/systemone endpoint on your own machine. Images work through the base model’s vision tower with no image training.
What it is good at
It is small and calibrated. The lab run on the Jev Decision Index (multimodalart’s community board on Hugging Face), which used hobson-v21, gives a calibration error of 0.0441, with mean confidence close to accuracy. It ran at a median 52.3ms per request there. The authors report 176 of 231 on the public JevBench tasks, and 180 for the newer v1; those runs are vendor-run, and the model has no row on the JevBench board.
What it is not for
Its accuracy is low on the broad suite. The Jev Decision Index gives it 21.75 overall and 25.41 on the public part, against 57.43 for Surogate Rune. Microsoft’s 54.8% covers only the 23 of 36 benchmarks it could answer.
The authors list the weak spots: long multi-step documents, Score and Noul questions on rubrics unlike the training mix, and questions, which it reads less closely than documents, so a reworded question often gets the same answer. The window is 4,096 tokens. Do not use its confidence to detect a missing image. The v21 checkpoint trained on 20,000 rows from datasets whose terms restrict commercial use, which the v1 release dropped, so check that before commercial use even though the weights are Apache 2.0.
Access
There is no hosted API or price. Install the strands-decider package from PyPI (version 0.1.0, 1 October 2026), or pull the weights from the StrandsAgents organisation on Hugging Face. The repo had 586 stars on 12 October 2026 and includes an example that gates an agent’s tool call on two yes/no decisions.
Specifications
| System One Score | 34.7 from 1 of 3 boardsJev Decision Index 34.7. How the score works |
| Question types | NoulChoiceScore |
| Max Choice options | Not documented |
| Score levels | Not documented |
| Questions per call | Not documented |
| Total context | 4,096 tokens |
| State budget | Not documented |
| Rate limit | None. It runs on your own hardware. |
| Endpoint | POST /v1/systemone on your own server (strands-decider serve) |
| Licence | Apache-2.0 |
| SDKs | Python: strands-decider |
The README says the models are trained and served at a 4,096-token window. A longer prompt is shortened by default: the state is cut and the question keeps its options. The --strict-window flag refuses such a prompt instead. The README says nothing caps how many options a question may carry, and the server encodes up to 32 questions per forward pass by default. The state-token budget is not published.
Versions
- StrandsAgents/strands-decider-2B-hobson-v21, 5 Oct 2026, The checkpoint the Jev Decision Index lab run used. A LoRA adapter on Qwen3.5-2B-Base plus a readout head, 1.9 billion parameters. It trained on 20,000 rows from datasets whose terms restrict commercial use, which the later v1 dropped. Release notes
- StrandsAgents/strands-decider-2B-qwen3.5-v1-2610, 9 Oct 2026, The README's current reference model: a soup of three training runs on a larger data mix. Also mirrored in Amazon's Hugging Face organisation. The authors score it 180 of 231 on the JevBench public set. Release notes
- StrandsAgents/strands-decider-2B-hobson-v19, 30 Sep 2026, The first public checkpoint. Release notes
- strands-labs/strands-decider, 29 Sep 2026, Code, training recipe and evaluations (Apache-2.0, 586 stars on 2026-10-12). The repo also lists four Gemma 4 models from 2B to 26B-A4B, released on 2026-10-09 and not covered on this page. Release notes
Use cases
What people use Strands-Decider 2B for, one page per pattern.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
Typed tool dispatch with System One models
Ask Jev one Choice question over your tool names and one per closed-set argument, so every value that comes back is a value the function already accepts. Nouls decide whether an optional argument was mentioned at all. Your code assembles the call and runs it.
LLM guardrails with System One models
Put one Jev request in front of an LLM and one behind it. Yes/no questions return the probability that each hazard holds, a Score rates how much harm complying would do, and your thresholds turn those numbers into pass, review, block, or a crisis path.
Agent routing and skill selection with System One models
An agent choosing from a long skill roster reads one truncated line per entry and often loads the wrong thing. Jev ranks every entry in one request and separately answers whether any skill applies at all, so the agent gets a short hint instead of a guess.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| Jev Decision Index 0.3.1 | Full score | 21.75 score | Independent | 12 Oct 2026 |
| Jev Decision Index 0.3.1 | Public score | 25.41 score | Independent | 12 Oct 2026 |
| Jev Decision Index 0.3.1 | Expected calibration error | 0.0441 ECE | Independent | 12 Oct 2026 |
| Microsoft 36-benchmark comparison | Accuracy | 54.8 % | vendor-run | 10 Oct 2026 |
| Strands Decider model card | JevBench public set, tasks right | 76.2 % | vendor-run | 5 Oct 2026 |
- Jev Decision Index, Full score: Row: Strands Decider 2B, 106th of 143 rows by this score. The row's note says it is a lab run of the authors' strands-decider serve (v21, 4,096-token window; prompts over the window rerun with the window lifted), with no submission from the authors. Counts 20% public benchmarks and 80% private tests; 0 is random guessing, 100 is perfect.
- Jev Decision Index, Public score: Row: Strands Decider 2B. Lab run, as above.
- Jev Decision Index, Expected calibration error: Row: Strands Decider 2B. Lower is better. Mean confidence was 0.576 against 0.541 accuracy, so it is well calibrated but not very accurate on this suite. Lab run, as above.
- Microsoft 36-benchmark comparison, Accuracy: Run by Microsoft on an undisclosed 36-benchmark suite to market its own model. This is the average over only the 23 of 36 benchmarks Strands-Decider 2B could answer, so it is not comparable with the full-suite means of other models. Microsoft labels the row AWS.
- Strands Decider model card, JevBench public set, tasks right: Run by the authors on the 231 public JevBench tasks for hobson-v21 at a 4,096-token window: 176 of 231, Brier 0.323, ECE 0.064. The authors' own page says 231 tasks are few and treats differences under about 10 tasks between single runs as unresolved. The model is not on the JevBench board.