The mysterious Hy3 LLM is topping OpenRouter Model Rankings by a large margin
17 hours ago
- OpenRouter provides a unified API for most LLMs and publishes usage data on its AI Model Rankings page, which is valuable due to the rapid release of new models.
- Hy3 preview, a free Chinese open-source model from Tencent, unexpectedly ranks high in token usage on OpenRouter, surpassing popular models like Claude by more than 50%.
- Despite its popularity, Hy3's quality is not competitive with top models like Claude Opus 4.7 or GPT 5.5, and its usage is organic but driven by paying users after a free period.
- The effective cost of LLM APIs is heavily influenced by prompt caching, with DeepSeek V4 Flash offering cheaper effective prices ($0.018/1M input tokens) than Hy3 ($0.034/1M) due to lower cache read costs.
- DeepSeek V4 Flash's low cost is tied to its provider-specific caching advantages, but concerns about data privacy with Chinese companies may limit its adoption.
- The author cannot explain Hy3's popularity and speculates it may be used by a single large app for data processing, not coding, while noting that DeepSeek V4 Flash could gain traction if users switch.