ChinaA state-owned telecom released a model that fills in command parameters.
It published the weights, training corpus and inference toolkit together, not the model alone.
One model classifies a spoken request and extracts values such as room and brightness into structured data, with latency under 150 milliseconds, the lab says.
The lab’s own benchmark puts its zero-shot Micro-F1 at 70.27, above GPT-4.1.
Zhidx’s 11 tests through the lab’s interface are the only outside check reported.
How each outlet framed it
- Zhidx leans favorable
- uses Jev's decision-vs-generation split to argue intent models still can't extract structured parameters, testing Xing4.0-29B-Intent on 11 smart-home and car scenarios
Sources: Zhidx