AWS releases open-source 2B decision model for agent workflow routing
AI agents · Friday, 2 October 2026
Why it matters
Agent builders can now test a lightweight, open decision checkpoint locally rather than relying on a proprietary hosted router, which could support lower-latency and more controllable tool-use pipelines. The tradeoff is operational: teams must provide their own compute, and the reported benchmark and latency results do not establish that self-hosting is cheaper or more accurate than TypeSafe’s hosted Jev.
What happened
Amazon Web Services released Strands Decider 2B, a 2-billion-parameter model that scores predefined choices and returns confidence distributions before an agent makes tool calls. AWS adapted Alibaba’s Qwen3.5-2B with a scoring head and rank-16 LoRA update, adding just over 1 million parameters; the model is available on Hugging Face under an Apache 2.0 license, with no AWS-hosted API or published per-token operating cost. AWS reported about 72% accuracy and a 0.35 Brier score on JevBench, while its local testing recorded a 106-millisecond median response on an Nvidia RTX 3090; the article says this trailed Mapika’s decider-2b v11.
Players & places
- Amazon Web Services
- Strands Labs
- Alibaba
- TypeSafe
- Mapika
- Marc Brooker