Strands Decider 2B: a small, open-source, decision model
Points and comments are a snapshot, not live.
Strands Decider 2B is a small, open-source decision model for fast, local agentic workflows.
Strands Decider 2B is a 2-billion-parameter decision model built on a Qwen3.5-2B torso with a pointer head replacing the LM head. It is designed to pick between options and assign confidence scores, not generate text. The model achieves median latency of 115ms on an RTX3090 and 153ms on an M3 MacBook, with competitive accuracy and calibration on JevBench. It is open source on GitHub and Hugging Face, including training data and scripts. The team envisions use cases including model routing, tool selection, evaluations, guardrails, and hybrid agentic workflows with LLMs.
What commenters are saying
Commenters were largely positive, praising clarity, local deployment, and speed. Several shared practical experiences: running it in a browser via WebGPU, on an M3 Mac CPU in ~1.7s, and in Docker. Some questioned the term "noul" (attributed to Bernoulli maps) and whether the model can run on NPUs. A few noted the model's option-order sensitivity. One commenter linked to the similar Intern-Decision family, and another highlighted Cloudflare's Clef as a multimodal alternative. The thread included light humor about human-backed decision models.