d3-ultra answers decision questions with probabilities

The model card describes a decision model: software supplies material and questions, and receives probabilities for possible answers. One request can combine choices, yes-or-no questions and visual locations. The input is read once. A location answer always contains one box, so its confidence matters: a near-zero value indicates that the requested object is absent. The developers’ evaluations use private tests, not an official leaderboard.

The runtime change, merged on October 25, adds serving support for d3-ultra. The previous runtime rejected its package because the declared model identity differed. The change handles its special text layout, placing questions before the supplied material; requests containing images or video retain the earlier layout. The authors also record a limitation: closely tied choices can produce different decisions from the model package’s own runtime.

For developers, the surrounding Semantic Router project clarifies where such decision software fits. Its router selects or combines configured models according to a policy. The application running the agent still owns its repeated work, tools and task state. Local, private and cloud backends are selectable destinations, with quality, response time and cost as configurable priorities. The project directs applications to its inference endpoint; routing does not transfer ownership of the agent’s work loop.

MarkTechPost’s coverage, by Asif Razzaq, summarizes the release. It is secondary coverage, rather than an independent test.