A local LLM runtime for running very large MoE models on consumer NVIDIA GPUs, with quantization caches and an OpenAI-compatible API.