China’s Open-Weight Bet

Ask which country leads in AI and the debate fixes on frontier models. It is the wrong scoreboard. The more consequential contest is over the open-weight layer, and there China is not competing, it is winning.

Alibaba’s Qwen is now the most-downloaded open model family in the world, having overtaken Meta’s Llama. Hugging Face’s own tally counted roughly 2.0 billion Qwen downloads and more than 151,000 Qwen-derived models on its hub.^[1] By mid-2025, Chinese models held the top open-source ranks on the LMArena leaderboard: Moonshot’s Kimi K2 first among open models, DeepSeek second, with MiniMax and Qwen close behind.^[2] Chinese developers accounted for more than 45% of top open-model downloads in 2025.^[3]

The base model is not the product; it is the standard

Qwen has become the default foundation of the open ecosystem: by August 2025, more than 40% of new language-model derivatives on Hugging Face were built on it, against roughly 15% on Llama.^[4] The adoption is bleeding into Western production stacks. As Andreessen Horowitz’s Martin Casado put it, among Valley startups running an open stack, “there’s about an 80% chance they’re running on Chinese open models.”^[4]

Give the model away and it stops being a product. It becomes infrastructure, and infrastructure is where standards are set.

Why give away the crown jewels? Because a permissive licence turns a model into embedded infrastructure. Developers worldwide can download, fine-tune and self-host for free, with no API gatekeeper, and the base they choose becomes a de facto standard, reinforced by aggressive pricing and by research published alongside the weights. (One caveat for the record: Alibaba’s own download claims run higher than Hugging Face’s independent count, because they fold in Chinese platforms such as ModelScope; we cite the independent figure.^[1])

The mechanism is worth being precise about. When a developer chooses a base model to fine-tune, they inherit its tokenizer, its quirks, its prompt conventions and its upgrade path; switching later is costly. So the model that wins the default position accrues a compounding advantage that looks less like a product win and more like a platform one. That is why the churn among Chinese labs at the top of the open leaderboards, Kimi, DeepSeek, MiniMax, Qwen trading places, matters less than the fact that the base layer they collectively occupy is Chinese. The competition is fierce; the aggregate outcome is a standard.

Where the value goes

The open-weight strategy means China’s influence over global AI will be felt less through flagship consumer apps than through an invisible layer of models embedded in everyone else’s software. For investors, the leaderboard is a distraction. The value accrues around the substrate, inference, fine-tuning, agent frameworks and vertical applications built on top of freely available weights. That is the layer we underwrite.