Amazon publishes Strands Decider 2B, a small decision model for agent workflows
A 2B LoRA on Qwen3.5-2B-Base, Apache-2.0, with a /v1/systemone server. The card reports 0.779 on 231 public JevBench tasks.
Amazon published strands-decider-2B-qwen3.5-v1-2610 on Hugging Face on 9 October 2026. The card calls it a small, fast decision model for agentic AI. It answers three question types: yes/no, one of N options, and a level on an ordered scale. Each answer carries a calibrated confidence. It is not a text generator.
It is about 2B parameters: a LoRA adapter on Qwen3.5-2B-Base plus a small readout head. The v1 checkpoint is a soup of three training runs. The licence is Apache-2.0. You install it with pip install strands-decider, or run a server that exposes /v1/systemone, the same route used by Jev-style models.
The card reports 180 of 231 correct (0.779) on the public JevBench set, Brier 0.280 and ECE 0.072. It reports 0.653 on 6,000 held-out short tasks.
The limits are listed plainly. Context is 4,096 tokens. Long multi-step documents are weak. Transfer to rubrics unlike the training mix is poor. Image support was not trained. The HTTP server has no authentication and binds to localhost, so it is for local experiments.
A small model with an honest limits section. Set thresholds on your own traffic, as the card says.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights, code and a pip package are public; the HTTP server is local-only.
- Useful 3.5: Routing, tool selection and guardrail checks inside agents; context is only 4,096 tokens.
- Notable 3.5: A hyperscaler shipping open decision weights aimed at agent loops.
- Novel 3: Follows the Jev API format; the agent-framework tie-in is the angle.
Comments