AWS launches Strands Decider 2B, a local decision model

AWS released Strands Decider 2B, a downloadable decision model that picks from developer-supplied options instead of generating text. It builds on Qwen3.5-2B with a pointer head of just over a million parameters and a rank-16 LoRA adapter. AWS reports decisions in under 100 ms on an RTX 3090 and about 150 ms median on an M3 MacBook.

• Built on Qwen3.5-2B with a 1M-parameter pointer head and rank-16 LoRA
• Under 100 ms per decision on RTX 3090, ~150 ms median on M3 MacBook
• Second among public ~2B-parameter models on JevBench
• Training data, scripts and all earlier iterations published in the repo