Skip to content- Sources: primary, discussion
- Summary: Strands Decider 2B is a 2 billion parameter decision model announced on 2026-10-01 and released with weights on Hugging Face plus the training data and scripts, which is a fuller release than most model drops, and it reaches this digest through a Hacker News thread that surfaced today rather than as a release of the day. The architecture takes a Qwen3.5-2B torso, removes the language modelling head and replaces it with a roughly one million parameter pointer head that scores option positions against an answer position, with the torso fine-tuned through a rank-16 LoRA adapter. AWS reports a median decision latency around 115ms on an RTX 3090 and about 153ms for small tasks on an M3 MacBook, plus placement 3rd of 33 in the 2B class on JevBench accuracy and Brier calibration.
- Why it matters: Alongside the OpenAI Decisions API the same week, this marks decision models as a distinct class aimed at the routing, tool-gating and guardrail decisions where a full LLM call is too slow and too expensive to sit in the path.