Chinese open-weight models are now dominating Hugging Face downloads by 1.6B over US models

Chinese open-weight models reached 3.2 billion total Hugging Face downloads as of August 2025, double the total for American models, according to the article's cited download figures, with Alibaba's Qwen series driving China's lead starting around July 2025. On the Artificial Analysis Intelligence Index as of September 14, 2026, Z.ai GLM-5.3 scored 45, Moonshot AI Kimi K3 scored 44, Z.ai GLM-5.3-Flash scored 42, and Thinking Machines Inkling/Inkling Small scored 26, ahead of US open-weight models. The article distinguishes open-weight releases such as Llama, Qwen, Gemma, and DeepSeek from true open-source models like the Allen Institute for AI's Olmo, OpenAthena's Marin, and EleutherAI's Pythia, and notes Nvidia's Nemotron releases substantial training data under permissive licenses without being fully open-source.

Chinese open-weight models are now dominating Hugging Face downloads by 1.6B over US models Most of the teams I work with are shifting their stack toward open-weight models because the gap between them and closed APIs has basically vanished. If you're still thinking of "open" as just a cheaper alternative to GPT-4, you're missing the current reality. Right now, Chinese open-weight models are leading the pack in both adoption and raw performance, specifically in agentic capabilities that we previously only saw in closed systems like Claude Code https://promptcube3.com/en/tags/claude%20code/ back in December 2025. What actually counts as "open" in 2026? There is a huge difference between open-weight and true open-source, and it matters for how we deploy these at work. Most of what we use—like Llama, Qwen, Gemma, or DeepSeek https://promptcube3.com/en/tags/deepseek/ —are open-weight. You get the weights and the inference code usually via VLLM, SGLANG, or Transformers , but you don't get the full recipe. True open-source is much rarer. To qualify, you need the weights, the licenses, the inference code, and the complete training data and code to reproduce the model from scratch. The US is actually stronger here with non-profit efforts like the Olmo models from the Allen Institute for AI, OpenAthena’s Marin, and EleutherAI’s Pythia. Everything else exists on a spectrum. Nvidia’s Nemotron models, for example, release a ton of training data under permissive licenses, making them "more open" than a standard open-weight model, but they still aren't fully open-source since not every single piece of data is public. The shift in performance and adoption The momentum shifted about 18 months ago. If you look at the Hugging Face download numbers, China took the lead around July 2025, largely driven by Alibaba’s Qwen series. As of August 2025, the gap has widened significantly. Total downloads for Chinese models hit 3.2B, which is double the total for American models. The technical lead is becoming harder to ignore on benchmarks like the Artificial Analysis Intelligence Index AAII . As of September 14, 2026, the top-tier Chinese models are hitting numbers that leave US open-weight models in the dust: - Z.ai GLM-5.3: 45 - Moonshot AI Kimi K3: 44 - Z.ai GLM-5.3-Flash: 42 - Thinking Machines Inkling / Inkling Small: 26 Claude https://promptcube3.com/en/tags/claude/ Code last year, meaning these open-weight options can actually handle complex, multi-step tasks that used to require a top-tier closed API. Next AI Careers Still Mean Managing Human Chaos, Not Escaping It → https://promptcube3.com/en/threads/9597/ All Replies （1） Want a live back-and-forth? Join the global AI chat room https://promptcube3.com/en/chat/ — login to talk. That 1.6B download lead is wild. Most people overlook how much the localized tokenization helps with speed in these models.