Strands Decider 2B
AWS Strands Labs' open-weight 2B decision model that picks among predefined options in roughly 100ms on a single consumer GPU.
Overview
AWS's Strands Labs team shipped Strands Decider 2B on October 1, 2026, a 1.9-billion-parameter open-weight model built on a Qwen3.5-2B backbone, joining Cloudflare's Clef/Clef-flash and TypeSafe's Jev in the emerging 'decision model' category. Rather than generating text, it reads a state and a set of typed questions and returns a choice, a yes/no probability, or a calibrated confidence score, with a reported median latency around 106-115ms on an Nvidia RTX 3090. It runs locally on CPU, a consumer GPU, or Apple silicon, and is published under Apache 2.0 with weights on Hugging Face (StrandsAgents account) and training scripts/code in the strands-labs/strands-decider GitHub repo.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 1.9B parameters on a Qwen3.5-2B backbone, runs on CPU, consumer GPU, or Apple silicon
- Returns a choice, yes/no probability, or calibrated confidence score instead of generated text
- Median latency of ~106-115ms on a single Nvidia RTX 3090
- Fully open: Apache 2.0 weights on Hugging Face, training scripts and docs on GitHub
- Third major open decision model to ship in roughly a week, after TypeSafe's Jev and Cloudflare's Clef/Clef-flash
Use cases
- Bounded in-agent-loop decisions: run a tool, choose a route, request approval
- Low-cost local inference for high-volume conditional logic
- On-premises or offline agent deployments needing no API calls
Pricing
Free, open-weight (Apache 2.0); self-hosted, so cost is just compute.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 1 Oct 2026
- Record updated
- 4 Oct 2026
- Record last checked
- 4 Oct 2026