The pitch in Chongqing on 22 October was not about a smarter chatbot. It was about a factory floor: a production line where a city department's problem goes in at one end and a working agent comes out the other — in under 20 days, down from 90.
That claim came attached to Mingjing, a platform launched at the 2026 General Artificial Intelligence Conference by the Chongqing Institute for General Artificial Intelligence. Dong Le, the institute's executive dean, presented it. Science and Technology Daily, reporting the launch on 23 October, described it not as another general-purpose model but as an “intelligent-agent production line” for city governance, industrial upgrading and public services. China News Service covered the same opening on 22 October with the same framing.
Worth pausing on the phrase. “Agent production line” is a manufacturing metaphor, and it is doing real rhetorical work: it shifts the argument away from model capability and towards throughput, cost and repeatability — the things a city procurement office actually cares about.
The architecture bet
The technical core is CUV — cognition, capability, value — attributed to the team of Zhu Songchun, who heads Peking University's School of Intelligence and chairs the Chongqing institute. The promoters frame it against mainstream large language models, which they call a “big data, small tasks” parrot paradigm. CUV is presented as a “small data, big tasks” crow paradigm: autonomous perception, cognition, reasoning and execution, driven by value rather than by data volume.
It is a genuinely interesting argument, and in the material supplied it is also an assertion. Neither report gives a benchmark, a baseline, an ablation or a head-to-head test against a large model. The claim that CUV addresses AI's problems with trust, explainability and autonomous evolution is the institute's own description of its own work. That does not make it false. It makes it unverified.
The deployment structure is easier to pin down: a “1-3-4-N” stack — one CUV core; three bases (a self-evolving agent base, a social simulation platform, a virtual-real closed-loop training ground); four service platforms split by user, named Tongjian for government, Tongyan for commercial, Tongjing for civil and Tongce for testing; and N real-world scenarios.
The numbers, and whose numbers they are
Three figures carried the launch. Industry-agent delivery time falls from 90 days to under 20, with comprehensive R&D efficiency up more than fourfold. The social simulation platform can drive a million agents in parallel, a city-level “digital sandbox” for rehearsing crowd flow, emergency evacuation and policy transmission. The training ground cuts a typical robot skill to five days with an 85% real-machine validation pass rate.
Read them as a vendor's datasheet, because that is what they are. There is no stated measurement protocol, no baseline condition, no sample definition and no third-party audit in either report. “A million agents” describes a capacity that can be simulated; it is not evidence of a million agents running against a live city. The 85% pass rate comes with no described test set. The fourfold efficiency gain comes with no described denominator.
None of that is unusual for a launch event. It is worth stating plainly, because the numbers will travel much further than the caveats.
The deployment claims
The reports also cite early results: disaster prevention shifted from post-event response to pre-event warning; in a public-sentiment project called “peace of mind”, coverage of residents' requests rose from 35% to 82%; R&D recruitment cycles fell about 60%; new-drug experiment cycles were compressed 23%; and the expected share of road-rage-related accidents was halved.
The 35%-to-82% jump is the most striking and the least defined. Coverage of what, measured how, over what period, across which residents, against what denominator? A rise from 35 to 82 is a large claim; without the measurement basis, it stays a slide.
Also announced: a “Mingjing Partner Build-Use-Synergy Plan” with companies in large models, embodied intelligence, computing power and chips, to co-build the platform and city applications.
The politics of the pitch
Zhu's framing is the part most likely to interest readers outside China. He positioned Chinese AGI as a public-infrastructure effort on the model of high-speed rail, road networks and the “village-to-village” programme — low cost, high efficiency, inclusive — explicitly not a race for a few firms' high profits and trillion-scale valuations, and not a plan to replace high-value human jobs. The Chongqing deployment, in this telling, is meant to be a replicable template rather than a single-city showcase.
That is a deliberate contrast with the venture-funded model, and it is also a claim that has to survive procurement. Public-infrastructure rhetoric and fourfold-efficiency metrics sit awkwardly together: one promises restraint, the other promises exactly the kind of gain that gets a budget approved.
What to watch
The test is not the next conference. It is whether a city department outside Chongqing can take delivery of an agent in 20 days, and whether anyone beyond the institute publishes the before-and-after numbers with a denominator attached.
Until then, Mingjing is best read as an argument about how AGI should be paid for and deployed — not as proof that it works. The production line has been announced. The factory audit has not.