Strands released Decider 2B weights, training code, data and evaluation records.
Strands releases a 2B open decision model for agent control
2026-10-01Verified announcementTRIAL
Strands released Decider 2B, an Apache-2.0 model that chooses among typed actions in one forward pass without text decoding. It combines a Qwen3.5-2B backbone, a small pointer head and LoRA adapters. Weights, training code, data and evaluation records are public; author tests report 72.3% accuracy and 115 ms median latency on an RTX 3090.
What this means for you
For agent builders, a compact decision model can replace general-purpose generation in routing, tool selection and guardrails. The public training and evaluation artifacts make latency, calibration and abstention independently testable in a day.
Evidence & related material
Technical details & scores
| Architecture | Qwen3.5-2B torso plus an approximately 1M-parameter pointer head and rank-16 LoRA |
|---|---|
| Parameters | 1.9B |
| Output | typed action selection in one forward pass; no text decoding |
| License | Apache-2.0 |
| Artifacts | weights, training code, datasets and research record |
| Author Results | JevBench 167/231 (72.3%); Brier 0.342; ECE 0.052 |
| Author Latency | median 115 ms on RTX 3090; 153 ms on M3 Pro |
Technical impact 3
Engineering value 4
Adoption signal 3
Maturity 3
Verification cost 2
Risk 2