Chinese Labs Now Ship 2.78T-Param Models While US Stalls Under 130B

The open model landscape has fundamentally flipped in 2026, and the numbers tell a stark story. Chinese labs now routinely ship frontier-scale models that dwarf anything American companies release on their own. China’s monthly ceiling ran between 754 billion and 2.78 trillion parameters. America’s own ceiling stayed under 130 billion in five of seven months, with only NVIDIA’s Nemotron 3 Ultra at 561B and Thinking Machines’ Inkling at 952B breaking the pattern.

What’s driving this? Moonshot, MiniMax, Xiaomi, and Z.ai have abandoned the old progression path of starting small. They publish almost nothing below 70B parameters. That strategy works now because building large is no longer a differentiator — Xiaomi and Meituan both cleared a trillion parameters this year, and neither was a household name in open weights twelve months ago. The community’s quantization layer makes massive models runnable within days, so labs don’t need small models to be reachable.

Meanwhile, the US has quietly shifted roles. AMD and NVIDIA each released more than 200 new model repositories this year, far ahead of everyone else. Hardware vendors have figured out that open models sell chips — a model optimized for your silicon and freely available is the clearest proof the hardware works. But most US releases above 100B parameters are conversions built on top of Chinese models, not original work. That’s a distribution and optimization layer, not model creation.

The raw count hides another truth: Hugging Face now hosts 2.96 million model repos, but 85.6% have fewer than 200 lifetime downloads. Just 1.5% of repositories account for 99.2% of all downloads. The attention economy and the adoption economy have fully diverged — of the top 25 models by downloads and top 25 by likes this year, exactly one appears on both lists. all-MiniLM-L6-v2 was pulled 1.55 billion times in seven months against just 5,156 likes, while Kimi-K3 got roughly 60 downloads per like. Likes measure excitement. Downloads measure dependency. They’re reading two different fields entirely.