--- title: "Chinese open-weight models quietly took over Hugging Face" date: 2026-11-09 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-open-weight-global-market language: en --- # Chinese open-weight models quietly took over Hugging Face > Qwen and DeepSeek became the most-downloaded open large models on Earth in 2025. Here is what the shift means for teams picking an AI stack. You open a notebook, type a command, and point it at a model someone shipped from Hangzhou or Beijing. A year ago that would have felt exotic. Today it is the default for a large slice of the world's developers. The large model (大模型) race is usually told as a story of who scores highest on a benchmark. There is a quieter story in the download logs, and it has already been decided. ## The download chart flipped in 2025 For most of the open era, Meta's Llama was the most-downloaded open large-model family on Hugging Face. That changed in September 2025. Alibaba's Qwen (通义千问) became the platform's most-downloaded model family, overtaking Llama. By mid-December 2025, Qwen had accumulated roughly 385 million downloads, while Llama sat at about 346 million, according to reporting by China National Radio (央广网) summarizing a Stanford HAI / DigiChina policy brief and The Rundown AI's tracking. The gap is not only about one family. Looking at the broader platform, models from Chinese developers accounted for 17.1% of all Hugging Face downloads between August 2024 and August 2025 — the first time that share passed the United States' 15.8%, per a Massachusetts Institute of Technology and Hugging Face dataset analyzed in the Stanford brief. ## What "open-weight" actually buys you Open-weight (开放权重) means the model's trained parameters are published. You can download them, run the model on your own hardware, and modify it without asking a vendor for permission or paying per token. Two Chinese releases define the category: - **DeepSeek-R1** (January 2025) ships under the MIT license with 671 billion total parameters and about 37 billion active per token via a Mixture-of-Experts (混合专家) design. It is a reasoning model (推理模型) trained largely by reinforcement learning, and it publishes its chain-of-thought (思维链) traces — something OpenAI's o1 does not. - **Qwen3** (April 29, 2025) uses the Apache 2.0 license, a 235-billion / 22-billion-active MoE architecture, and supports 119 languages and dialects. Both let a company self-host for data-sovereignty and cost reasons, and adopters have reported meaningful cost savings from running open-weight models in-house rather than paying per-token API fees for proprietary systems, though the size of the saving varies widely by workload and scale. ## Three things the chart does not show - **Downloads are not deployments.** The same model pulled ten times by one researcher counts ten times. Derivative models matter more: Chinese-derived models made up 63% of new derivative models published on Hugging Face in September 2025. - **"Open" is not one thing.** Some releases carry use restrictions; MIT and Apache 2.0 are the permissive end. Always read the license before you build a product on a weight. - **The US responded.** The White House's "America's AI Action Plan" in July 2025 elevated open-weight models to a strategic asset, and OpenAI later shipped an Apache-2.0 open-weight release — a direct response to competitive pressure, per the same reporting. ## Where the weights actually run Open-weight (开放权重) does not mean free to operate. The 671-billion-parameter DeepSeek-R1 needs roughly 1.3 terabytes of memory across GPUs to load at full precision, so most teams run the smaller distilled variants (1.5B to 70B) on a single machine. Qwen3's 235B flagship is similarly heavy; its 22B-active design keeps inference cheap once loaded. The practical divide is between "download and tinker" (easy) and "serve at production scale" (a real infrastructure project). That nuance gets lost in the download charts: a 385-million download count includes countless student experiments and one-line tests, not 385 million production deployments. It measures mindshare, not revenue. ## Why this matters outside China For developers in Southeast Asia, the Middle East, and Latin America, the practical effect is simple: a capable model no longer requires a US API key, a foreign credit card, or sending data to an offshore server. That is why Singapore's government AI program and Malaysia's sovereign AI plans both built on Chinese open-weight foundations. ## What readers can do now - **Prototype locally first.** If your task is sensitive (health, finance, legal), download an Apache-2.0 or MIT model like Qwen3 or DeepSeek-R1 and run it on hardware you control before paying for any API. - **Check the license, not the headline.** Permissive (MIT / Apache 2.0) lets you ship commercially; "open-weight" with extra clauses may not. - **Benchmark on your own data.** A leaderboard rank is a starting point, not a purchase order. Run a small internal eval before committing. ## Honest limitations This article leans on download statistics reported by Chinese state media and summarized from a Stanford policy brief; I did not independently audit Hugging Face's raw counters. "Downloads" measure attention and ecosystem activity, not revenue or production usage, and the figures shift weekly. I have not verified every vendor's cost-saving claim with the underlying companies. The open-versus-closed strategic framing reflects 2025 policy reporting and may move fast. ## Key takeaways - Qwen overtook Llama as Hugging Face's most-downloaded model family in September 2025, with roughly 385M vs 346M cumulative downloads. - Chinese developers' share of platform downloads (17.1%) first exceeded the US (15.8%) in the year to August 2025. - Open-weight, permissively licensed models let teams self-host for sovereignty and cost — but license terms and your own eval still decide the build. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-open-weight-global-market Free to quote with attribution and a link to the original.