Strands Decider, in your browser

A decision model (Qwen3.5 hybrid + pointer head) running entirely on your GPU: routing, intent, tool choice, agent checks, in many languages. No server inference, no API calls. Edit anything and it re-decides as you type.

loading…

State: what the model sees Question Options

Re-decide as you type

Decision

…

forward – tokens – decisions 0

Weights: an unofficial compressed build of StrandsAgents/strands-decider-2B-hobson-v21 (its LoRA merged into Qwen/Qwen3.5-2B-Base): decoder layers 14–22 replaced by one trained linear block, int4 GPTQ on Hadamard-rotated weights, and an int3 embedding whose 32k most common rows ship up front while the rest load on first use; 8-bit scales and gzip for the download. About 470 MB, cached in the browser's private storage (OPFS) after the first load. Backend: .