IonRouter

IonRouter is a managed inference API by Cumulus Labs that provides OpenAI-compatible access to open-source AI models at roughly half the cost of competitors. Powered by a custom IonAttention engine optimized for NVIDIA Grace Hopper hardware, it supports LLMs (Qwen, DeepSeek, GLM), vision, video generation, and text-to-speech models. Developers swap their base URL and get sub-100ms model swap times with up to 7,167 tokens/second throughput.