Chinese technology outlet 智东西 tested Xing4.0-29B-Intent in 11 groups of home, vehicle and office scenarios. Its hands-on assessment found examples of correctly matching objects and values, but an instruction forbidding rear-window opening was interpreted as a request to shut them. The single newsroom assessment examines a model that returns intent categories and parameters for downstream software.
Chinese technology newsroom 智东西 tested Xing4.0-29B-Intent, a model from China Telecom’s Xingchen general-AI laboratory, in 11 groups of scenarios. The single newsroom’s hands-on assessment found a problem in a vehicle-control example: when instructed to leave rear windows unopened, it returned zero as their opening value. If the windows were already partly open, executing that result would close them and change their existing state.[1]
The model combines identifying an intended operation with extracting the information needed to carry it out. Its JSON output, a structured data format used by software, contains category for the operation type and args for associated parameters. The assessment also covered home devices, work orders, expenses, meetings and messages. Successful examples matched different people or objects to their respective values rather than mixing them.[1]
In a note-taking example, wording about cancelling a meeting and sending a message stayed within the note. The model did not turn those words into extra tasks. Xing was fine-tuned from a base with 29 billion parameters. The training material was reorganized from 61 public datasets and covers 19 domains as well as 52 languages. Alongside the model, developers can obtain the data and tools for running inference. The assessment’s evidence comes from its limited series of scenarios, rather than independently replicated comparisons against other models.[1]
