Megvii

Contribution to AI

Giving the technology away first turned out to be the business model. Face++ opened as a developer API in 2012 and remained free to most users through 2013, which bought Megvii a distribution channel and a stream of training images no laboratory could have collected by asking; publish an interface, gather the data, sell the accuracy back as authentication became a familiar sequence among Chinese vision firms afterwards. The research that followed was not incidental to the product. Sun Jian arrived from Microsoft Research Asia in 2016, having co-authored the residual networks paper, and assembled a group that took first place in the COCO 2017 detection challenge with MegDet and published ShuffleNet, whose pointwise group convolutions and channel-shuffle operation kept classification usable at budgets of tens of megaflops — phone silicon rather than server racks. Objects365 and CrowdHuman, prepared with the Beijing Academy of Artificial Intelligence, gave detection work benchmarks for crowded scenes and long-tailed categories, and YOLOX later put an anchor-free detector into wide industrial use. MegEngine, built in-house from 2014 and open-sourced in March 2020, was an argument that a training stack need not be borrowed from California. The remaining legacy is regulatory: the October 2019 United States entity listing, imposed on Megvii alongside seven other Chinese firms over surveillance in Xinjiang, made export control a live constraint on computer vision and diverted the planned Hong Kong flotation to Shanghai.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).