Vivix Releases A1: First Real-Time Model for Interactive Virtual Character Communication
Vivix
Vivix has unveiled A1, the first real-time multimodal interactive model that enables users to have video-like conversations with virtual characters. The model features nearly 30B activated parameters, achieves over 10,000 video tokens per second on a single GPU, and runs on consumer-grade GPUs through innovations like MJD distillation and the VMI inference infrastructure. Additionally, Vivix introduced W1, which allows users to intervene and change storylines in real time.
Vivix released A1, a real-time interactive multimodal model that allows users to communicate with virtual characters via live video. Unlike traditional digital humans, A1 characters possess physical bodies that can move, change posture, and interact with their environment. The model uses a unified streaming architecture that integrates multimodal references, real-time interaction, and streaming generation. It features nearly 30 billion activated parameters and achieves over 10,000 video tokens per second on a single consumer-grade GPU, thanks to technologies like Multidimensional Joint Distillation (MJD) and the Vivix Model Infra (VMI). VMI decouples encoder and decoder, uses native NVFP4 training-inference co-design, and employs a compute-communication-memory co-scheduling engine to minimize latency. The system responds to new inputs in as little as 300 milliseconds, with an average delay of 0.6 seconds. Vivix also introduced W1, which lets users intervene and alter character choices and storylines in real time.
- Сокращения
- MJD = Multidimensional Joint Distillation — Многомерная совместная дистилляция
- VMI = Vivix Model Infra — Инфраструктура моделей Vivix
- NVFP4 = NVIDIA FP4 (4-bit floating point) — NVIDIA FP4 (4-битное число с плавающей точкой)
- GPU = Graphics Processing Unit — Графический процессор
- PD = Prefill/Decode — Prefill/Decode
- KV Cache = Key-Value Cache — Кэш ключ-значение
Source: QbitAI 量子位 —
original