MoshiRAG is released as the world’s first full-duplex speech language model with retrieval-augmented generation (RAG)! Check out the paper, code, and blog post. Excited to present it at ICML 2026!