Strands Decider 2B v1
Strands Decider 2B v1 scores typed questions with an open adapter and head. Read its runtime, three-run weight average and public benchmark results.
By Strands Labs · Hugging Face weight release: 9 Oct 2026
A LoRA adapter and option-scoring head on Qwen3.5-2B-Base for Choice, Noul and Score questions. This v1 release averages three training runs and supplies a System One server. Its optional vision path keeps the base vision tower without image training. The community SageMaker project deploys the separate Strands Decider 2B hobson-v19 checkpoint.
Quick answers
What is Strands Decider 2B v1 by Strands Labs?
Strands Decider 2B v1: A LoRA adapter and option-scoring head on Qwen3.5-2B-Base for Choice, Noul and Score questions.
Who makes Strands Decider 2B v1?
Camera test and hosting estimates
The camera-still report uses the existing v1 checkpoint, without new image training. Its 768 px Python path scores 85.2% at a test-selected cutoff. The HTTP path resizes to 448 px. Hosting costs depend on sustained volume; the reported L4 break-even is arithmetic without a measured L4 throughput run.
Camera evidence · Hosting utilisation example
A separate E4B production deployment
Amazon’s ticket-labelling report uses a different Gemma 4 E4B checkpoint, not this 2B v1 model. The team describes SageMaker deployment with a shadow-comparison mode and a 45-second configuration rollback. Its reported cost depends on busy instances and shorter prompts; the published account supplies no raw tickets or held-out accuracy total.
Production report and cost limits
Facts
- Base
- Qwen/Qwen3.5-2B-Base; separate base download required
- Inputs
- Text; optional image path needs the vision extra and a runtime release after 0.1.0
- Runtime
- strands-decider Python package, CPU, MPS or CUDA; POST /v1/systemone
- Licence
- Apache-2.0; LICENSE.md included
Claims and sources
- The maker reports 180/231 public JevBench tasks correct at a 4,096-token serving window, Brier 0.280 and ECE 0.072. CPU and NVIDIA L4 select the same answers on all 231 tasks; the training harness differs on one near tie. These are maker runs, not an independently reproduced board score. vendor reportedsource
Links
Source: Card, adapter and run records → Source: Licence → Source: Runtime and training code →
Related projects
No related projects are listed in the registry.