Griffin: The First Human Interaction Model

Griffin is a new class of Human Interaction Model (HIM) that performs real-time, face-to-face conversation by perceiving audio and video continuously and generating synchronized speech, facial expressions, gestures, and whole-scene video pixels on the fly. It replaces cascaded pipelines (speech-to-text → LM → renderer) with a unified video-to-video, full‑duplex system: a continuous conversational modeling engine decides when and how to respond every sub-second, while streaming speech and video generators convert those decisions into audiovisual behavior. In live studies Griffin achieved a 48% pass rate on a real-time video Turing test (nearly half of participants thought they spoke with a real person), ranked first on an independent NVIDIA benchmark, and outperformed the next best system by 37%. A limited Griffin‑Lite preview is available to early testers, with broader model releases planned.

Capabilities emphasize natural, timed interaction: back‑channeling, interrupting or waiting appropriately, emotion and behavior modeling, temporal awareness of silences, and perception-driven responses (e.g., coaching during tasks, reacting to objects or gestures). It generates full scenes - face, body, shadows and background - from a reference image in real time. The technical core is concurrent perception, decision-making, and generation so the model can alter its output mid‑utterance. Safety, limitations, and responsible deployment practices are acknowledged, and further evaluation and expansion are forthcoming.