Start with the counting rule. A repository on the Hub is one address for one model [13]. Two hundred of them says nothing about how many were trained rather than repacked, and the dev.to summary of the report concedes the point in its own closing caveats: the chip vendors post in volume partly because much of the output is existing models converted to run on their hardware [10].
Set that against the derivative flow. Qwen derivatives grew at 180 to 210 new repositories a day across the first seven months of the year [4]. At roughly 210 days, that is between 37,800 and 44,100 repositories built on one family [1]. The 400-plus first-party repositories from AMD and NVIDIA together amount to about two days of that rate [2]. The ranking measures first-party publishing throughput, and ecosystem gravity is a separate column in the same report.
The reason given for hardware vendors publishing at all is blunt. A tuned model that runs free on your silicon is the strongest available proof the silicon works [2]. AMD is described as a distribution and improvement layer, its conversion work getting trillion-parameter Chinese models onto Western hardware, while Chinese labs run the same play in reverse and tune for domestic chips [3]. Two hundred repositories is a marketing budget expressed in git pushes.
What the material does not say is what a conversion contains. Format, quantisation, kernel target and runtime version are the parts a buyer needs, because a converted checkpoint only claims that this model, at whatever precision the vendor chose, ran on the vendor's stack. Your context lengths and batch sizes are not inside that claim. It is a starting configuration you did not have to build, not a result you can quote.
The licence movement is where the hardware reading sharpens. Among 178 Chinese models above 20B released this year, 59 percent shipped under Apache 2.0 and 22 percent under MIT, with almost no commercial restrictions [7]. In recent weeks Kimi K3 and Qwen 3.8 2.4T attached non-commercial limits and revenue-share terms, K3's disclosed in its own technical report [8]. The report's cautious conclusion is that when free weights are not earning licence revenue, the return has to arrive through an API business, a hardware position or an ecosystem position [9]. A chip vendor is the one publisher on that list with no reason ever to tighten a licence, because the return is booked when the buyer picks the accelerator.
One caution on the evidence. All of it reaches me through a single secondary summary, published on dev.to on 4 September 2026, written by a model called glm-5.3 under human review, of a Hugging Face report I have not read [12]. Its author also warns that download counters can double-count enterprise automation pipelines [11]. That is enough to support a direction. The 55x download gap between Qwen and Moonshot, 2,045 million against 37 million, survives noisy counters because the ratio is large [5][3]. A 200-against-200 ordering between two vendors would not.
