Strands Decider 2B is a free, open-source decision model from AWS's Strands Labs, built for AI-agent developers who need cheap, fast, structured choices inside an agent loop. Amazon released it on October 1, 2026.
Core Features
- No text generation: instead of writing answers, it picks from predefined options or scores a choice, returning a calibrated confidence value in a single forward pass.
- Built on Qwen3.5-2B: AWS replaced the language head with a small pointer head and added a rank-16 LoRA, keeping the total at about 1.9 billion parameters.
- Runs locally: it executes on CPU, consumer GPU, or Apple silicon with no internet needed after download, and no hosted API or per-call fee.
- Three question types: choice, noul, and score, all constrained to allowed outputs.
- Apache-2.0 weights, code, training data, and recipe published on Hugging Face, with a bundled CLI and local HTTP server.
Use Cases / Best For
- Agent routing and tool selection, where a frontier model would otherwise bill for a yes or no.
- Guardrails and policy classification that must stay fast and on-premises.
- Teams that want to fine-tune their own routing classifier on owned hardware.
Pros / Cons
- Pro: Free Apache-2.0 weights you can self-host with no per-call fee.
- Pro: 115ms median latency on a single GPU for bounded decisions.
- Con: Cannot write text, so it is useless for chat, code, or summaries.
- Con: You must run and secure the local server yourself.
Pricing
Strands Decider 2B is free under Apache-2.0 ($0 to use). There is no hosted endpoint on Bedrock or any inference provider, so you pay only for the hardware you run it on. AWS measured a 115-millisecond median latency on an RTX 3090 and about 153 milliseconds on an M3 Pro. It ranked third of 33 models in its size class on JevBench v1.4.2.
Our Take: Best for builders who want a self-hosted, auditable gate inside agent pipelines; the trade-off is that it cannot code, chat, or summarize, so pair it with a reasoning model for hard calls.
Find more AI Models and Engines.
