中国前沿模型规模领先，AMD与NVIDIA主导发布量

Key Highlights

A Hugging Face blog revealed thought-provoking data: from January to August 2026, public model repositories on the platform grew from 2.43 million to 2.96 million, still surging in number, yet 85.6% of models were downloaded fewer than 200 times, and only 1.5% of repos captured 99.2% of downloads. At the same time, Chinese labs' largest monthly open-source model ranged from 754B to 2.78 trillion parameters, while US labs stayed below 130B in five of seven months. In one sentence: the more models, the more crowded, yet the head-effect becomes more extreme, because abundance does not mean attention is spread evenly. The numbers paint a market where publishing is cheap but being used is rare, and where the headline contest over parameter scale is led by one region while the headline contest over release volume is led by hardware vendors. Both facts together describe an ecosystem that is simultaneously more open and more concentrated than it looks.

What It Does and How It Unfolds

The report also points out a contrast: the leaders in release volume are not model companies but chip vendors. AMD and NVIDIA each released over 200 new model repos, becoming the most prolific open-source publishers, a role many would have expected a lab like Meta or a Chinese giant to hold. The logic is clear, those who sell compute have the strongest incentive to help developers use their chips via companion models that are already tuned and benchmarked for the hardware. Chinese labs take a few but large route, using a handful of super-large models to seize the high ground of parameter scale and grab headlines; US labs are more restrained in most months, with generally smaller parameters focused on vertical and practical use rather than spectacle. The divergence shows two strategies for influence: one seeks it through size, the other through fit, while the chip makers quietly win on sheer count of things shipped.

Technical Details

Highly concentrated downloads show that attention in the open-source community is extremely scarce: of the vast number of models, only a tiny fraction are actually used, and the rest form a long tail that almost nobody touches. The 1.5% of repos eating 99.2% of downloads is nearly a power-law distribution, the same shape seen in many online ecosystems where the winner takes nearly everything. The Chinese models' 754B to 2.78T range reflects the spread of the MoE architecture, where total parameters can be huge while activated parameters stay controlled, balancing scale show with actual usability in a single design. AMD and NVIDIA releasing many models is more about optimizing and demoing around their own hardware, an ecosystem play rather than pure research output, so the models should be read as documentation with weights attached. Understanding this distinction prevents the mistaken conclusion that chip vendors are suddenly leading AI research; they are leading AI adoption plumbing, which is a different and very deliberate game.

Comparison With Competitors

Putting China and the US together: Chinese labs prefer charging parameter scale, using super-large open-source models to brush presence and influence on leaderboards and in the news. US labs except chip vendors lean toward smaller scale emphasizing usability and compliance, which matters more for enterprise adoption where risk and cost dominate. On the chip side, AMD and NVIDIA do not compete on single-model size but on the breadth of the model matrix, binding their hardware ecosystem with model libraries covering various tasks from inference to fine-tuning. Put simply, one side compares who is biggest, another compares who is most complete, and chip vendors compare who best helps you use it, three different scoreboards for the same sport. The result is a richer but more confusing landscape where a developer must decide whether they want a trophy model, a dependable tool, or a hardware-matched accelerator, because no single publisher leads on all three at once.

Industry Impact and Use Cases

For developers, this data is a reminder: do not be fooled by the number of repos; what truly deserves attention is still that 1.5% of high-download models that the community has actually validated through use. For domestic labs, super-large models are a voice weapon, but download concentration hints that released does not mean used, and usability, docs, and ecosystem need follow-up work or the investment yields only press rather than adoption. For chip vendors entering model publishing themselves, it signals that hardware plus model integrated competition will intensify, as the boundary between selling silicon and selling software keeps blurring. Beneath the prosperity of the open-source world, the Matthew effect remains cold, rewarding the already-popular and punishing the obscure regardless of intrinsic quality. The practical takeaway is to build for usage, not for the leaderboard, because in this market attention is the scarce resource, not models, and the labs that internalize this will outlast those that only chase parameter records.