The LLM that came out of nowhere: How Hy3 climbed to the top of the charts

A largely unknown AI model called Hy3 preview has shot to the top of OpenRouter’s AI Model Rankings, outpacing even the widely used Claude by more than 50% in token usage. Max Woolf writes in his blog that the rise is puzzling, because the model’s quality does not appear to justify its popularity.

Hy3 preview is an open-source model released by Chinese technology company Tencent. Its benchmark results, published on Hugging Face, are notably modest and do not compare favourably to other Chinese open-source models. Despite this, the model now ranks above established names like Claude and DeepSeek V4 Flash on OpenRouter, a widely used service that routes API requests to various AI providers.

Usage pattern raises questions

Hy3 preview was initially available for free on OpenRouter starting around May 8. When it moved to a paid model, usage did not drop significantly, which Woolf notes is unusual. If users had not been getting value from the free version, he argues, they would likely have stopped once costs kicked in.

The model is served exclusively through a single provider, the Singapore-based SiliconFlow. It is priced at $0.066 per million input tokens. Woolf points out that this is not the cheapest option available: DeepSeek V4 Flash, when served directly by DeepSeek, has an effective price of around $0.018 per million input tokens due to aggressive prompt caching.

Usage data shows that individual apps account for only a small share of Hy3’s traffic. Woolf’s best explanation is that a single large application, not affiliated with Tencent, is using Hy3 as its data-processing backbone. The identity of that application remains unknown.

Hy3 preview’s rise is not linked to any obvious marketing push. There is almost no discussion of the model on Hacker News and only limited coverage on Reddit. Woolf concludes that the rankings mystery remains unsolved.