AWS reports Strands Decider 2B decisions under 100 milliseconds on RTX 3090

AWS also reports a roughly 150-millisecond median response time for small tasks on an M3 MacBook. The model scores supplied options rather than generating text.

AWS also reports a roughly 150-millisecond median response time for small tasks on an M3 MacBook. The model scores supplied options rather than generating text.
AWS reports that Strands Decider 2B makes decisions in under 100 milliseconds on an Nvidia RTX 3090 graphics card, with response times increasing as tasks grow. On an M3 MacBook, the company reports a median response time of around 150 milliseconds for small tasks. These are company-reported measurements, not independent test results. Detailed test conditions are not specified alongside the figures, limiting comparisons between devices and workloads.
Amazon Web Services has released Strands Decider 2B, an open-source superintelligence (SI) model that chooses among predefined options instead of generating text. According to TechCrunch’s October 1 report, it is available now and small enough to run locally. Its intended role is narrow: deciding what an SI workflow should do next without using a full text-generating model for every step.
Strands Decider uses Qwen3.5-2B, a large language model, as its language-understanding backbone. As The New Stack explains, AWS removed the component that generates text and replaced it with a “pointer head”: a component that scores the answer options supplied by the developer. The model then selects an option and returns a confidence score rather than an open-ended written response. Developers can use these choices to route requests, select tools or check proposed actions, while leaving conversation and more complex work to text-generating models.
Marc Brooker, identified by TechCrunch as an Amazon distinguished engineer, said the idea grew from conversations with AWS customers. Their SI-agent workflows—sequences of tasks handled by SI software—did not always need the capabilities or cost of a fully featured language model. He described the decision model’s role as choosing the next action based on the workflow’s current state.
Brooker said confidence scores and the restricted set of possible answers could make a workflow step more reliable, with faster responses and potentially lower costs. Those are AWS’s claims. Restricting the answers prevents the model from inventing an option that was not supplied, but does not guarantee that it selects the correct answer.
Brooker began the project after seeing Jev, a decision model from TypeSafe. TechCrunch reports that his project briefly reached first place on the Jevbench ranking for models of its size before Amazon engineers prepared it for release through Strands Labs, which develops tools and protocols for deploying SI agents. That benchmark placing does not establish dependable performance across real workflows.
Brooker also described a development trade-off: improving decision accuracy and calibration—how well confidence scores match the correctness of decisions—without weakening the model’s understanding of different languages or its general knowledge. The challenge is not simply to make choices faster, but to preserve the understanding needed to make useful choices.

Proofpoint linked the campaigns, chiefly in July 2026, to a group it describes as China-aligned. The phishing pages could intercept passwords and tokens that keep users signed in.
2 min

OpenAI stands by the dismissals and says contracts with outside safety assessors are being finalised. The researchers allege retaliation, which the company denies.
4 min

IMPEL and IMPF want songs generally valued at least as highly as recordings, with payments and disclosure covering past use, SI training and generated music.