strands-decider-2B-hobson-v21
A 2B decision model from Strands Labs, the experimental arm of the Strands Agents SDK. A Qwen3.5-2B-Base torso with a rank-16 LoRA and a pointer head of about a million parameters answers noul, choice and score questions in one forward pass, with a probability per option.
The maker calls it experimental. The LM head is dropped, so it cannot generate text; option sets come from the request. The repo holds the LoRA adapter and head only, and the base weights download from Qwen at first use. pip install strands-decider gives an ask CLI and a local server with a /v1/systemone endpoint (no authentication). Calibration is one temperature per question type, fitted on short classification tasks; long multi-step documents are its weak spot (0.550 on the hard tier), and score and noul transfer poorly to rubrics unlike the training mix. The optional vision mode uses the base model's untrained vision tower, and its confidence does not drop when an image is missing. JevBench is a third-party benchmark; the figures are the maker's own run (176/231, Brier 0.323; six-seed mean 175.0). Latency was measured on v19, which stays published as the earlier checkpoint. The training datasets carry their own licences (data/sources.md in the code repo).
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 1.9B |
| Base model | Qwen/Qwen3.5-2B-Base |
| Maker | Strands Agents (Strands Labs) |
| Released | 2026-10-05 |
| License | apache-2.0 |
| Reported accuracy | 76.2% |
| Reported latency | 115 ms p50 / 299 ms p95 per question on an RTX 3090 under WSL2 (measured on v19, same architecture and size); 153 ms warm median on an M3 Pro |
Get the weights
pip install systemonemodels
systemone pull strands-agents/strands-decider
The files are served from the maker's Hugging Face repository, StrandsAgents/strands-decider-2B-hobson-v21, and verified against the checksums recorded here.