Qwen3.8 27B Addition-in-Words Test on Local DGX Spark Hardware

It tested whether the model could add large numbers and then write the result in words rather than digits, a task that reveals whether the model actually computes or relies on language patterns.

With reasoning disabled, errors appeared on larger numbers. With reasoning enabled, the model succeeded in 167 of 169 attempts, though each combination was tested only once.

The result is encouraging for running mid-sized models locally on simple arithmetic tasks, but it is a single experiment on a single hardware setup and is not enough for production adoption without internal testing.

Local execution gives full control over the environment and keeps data private, which matters for government and regulated organizations in the region that require in-house data processing.