Production Use, Not Hype, Drives Open-Model Adoption

Hugging Face's biannual landscape report, published August 14, shows how the open-weights ecosystem evolved from January through August 2026. The distribution is extreme: 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of all downloads.

The Hub grew to 2.96 million model repositories (up from 2.43 million at the start of the year), 1 million datasets, and 1.44 million Spaces.

China's research labs — Moonshot, MiniMax, Xiaomi, Z.ai — skipped the small-to-large progression that defined earlier open-model strategies and deployed frontier-scale models directly. China's monthly parameter ceiling ran between 754B and 2.78 trillion in 2026; the U.S. ceiling stayed below 130B in five of seven months. NVIDIA's Nemotron 3 Ultra (561B) and Thinking Machines Lab's Inkling (952B) were the primary U.S. exceptions above 100B. Most other U.S. releases above 100B are ports or conversions built on Chinese base models.

Two forces enabled the shift. First, the community quantization layer: a trillion-parameter release becomes runnable within days, removing the practical barrier that once forced labs to ship smaller models first. Second, large scale stopped being a cost differentiator. Xiaomi and Meituan both cleared a trillion parameters in 2026 without significant open-weights presence a year prior. Portfolio strategy signals intent: a frontier-only portfolio bets on benchmark position and API demand; a full-spectrum portfolio — Tencent and Alibaba Qwen cover everything from under 1B to above 70B — bets on becoming the standardized family.

Within U.S. institutions, the top publishers by new repository count are hardware vendors. AMD and NVIDIA each released more than 200 new model repositories in 2026, far ahead of any model lab; LiquidAI ranked third at roughly 100. Google and Meta now rank below NVIDIA in new model releases. Meta has shifted toward closed flagship models. Open source has migrated from model labs to hardware and infrastructure companies, with freely available models serving as proof that the underlying silicon works.

The attention-versus-adoption split defines what matters. The top 25 repositories by 2026 downloads and the top 25 by likes share exactly one entry. Not a single model published in 2026 appears in the downloads top 25; thirteen of those 25 date from 2022. all-MiniLM-L6-v2 was pulled 1.55 billion times in seven months against 5,156 likes. Kimi-K3 was pulled about 60 times per like received. Likes record what the field finds significant in the weeks after release; downloads record what is actually running in production.

For architects evaluating migration paths: the quantization layer makes most of the frontier technically accessible, but accessibility is not production readiness. The 85.6% of repos with sub-200 downloads is noise. Treat likes as an early-warning signal for what to evaluate, downloads as the signal for what has been validated at scale.