State of Open Models: Summer 2026 Observations

Hugging Face's mid-2026 analysis reveals a shifting open-model landscape. Chinese labs now lead the frontier, with monthly model sizes between 754B and 2.78 trillion parameters, while U.S. releases above 100B are rare, with only Thinking Machines' Inkling (952B), NVIDIA's Nemotron 3 Ultra (561B), Nemotron 3 Super (124B), and Arcee AI's Trinity-Large (399B) as major originals. AMD and NVIDIA each released over 200 new model repositories, far ahead of others, as open models serve as hardware marketing. Attention and adoption diverge sharply: only one repository appears in both the top 25 by downloads and top 25 by likes. Chinese labs license their largest models permissively, with 59% of releases above 20B parameters under Apache 2.0, likely to drive API and cloud business. Qwen has become the community's base model, with 151,448 derivatives, 2.6× Meta's footprint, and about 180-210 new derivatives daily. Small models under 1B parameters take 83% of all-time downloads, while llama.cpp enables trillion-parameter models like DeepSeek-V4-Flash and Kimi-K3 to run on consumer hardware, with Qwen-based GGUF downloads at 39.6 million monthly, nearly twice Gemma's and over five times Llama's.