Turpan, an oasis city in China's Xinjiang region, has quietly pulled off something its peers are still debating. The city's converged media center said on Nov 12 that its local government data center is now running two full-size models entirely on domestic computing hardware, the DeepSeek V4.1 Flash (552B) and the dense Qwen3-72B, serving internal government work. It's the first complete example of this kind in Xinjiang's government AI programs.
The build was finished on Nov 5 in a high-low pairing layout. The hardware is modest by data-center standards, just 2 domestic servers with 16 accelerator cards, yet it delivers 10 PFLOPS of compute and 1.5TB of unified memory at production grade. Data never leaves the government intranet. This is the first time Turpan has run models of mainstream open-source scale entirely on-premises.
Usage is already substantial. The government agent platform covers 51 municipal departments with 304 accounts, handling more than 500 office tasks a day across document proofreading, policy Q&A, meeting minutes and data summaries. One detail from the civil affairs bureau says a lot. Proofreading a public-welfare information report used to take several people most of a day; now it takes minutes.
The project sits under Turpan's "Digital-Smart Turpan" action plan. Local officials outlined the next steps, which include a municipal government AI validation center, multimodal upgrades, and gradual expansion to all departments and counties under a "mature one, promote one" approach. Pilot experience is expected to feed longer-term work on computing networks and compute-power coordination.
What Turpan demonstrates is simple. Full-stack domestic models can run end to end at the county level on a replicable, low-cost path. For peer governments, the barrier is no longer model capability but sustained investment in operations staff and scenario operations.