Vivix A1 Brings Real-Time Multimodal Interaction to Virtual Characters

Introduction

Vivix has released A1, a real-time interactive multimodal model, together with W1, a model aimed at user-steerable story worlds. The key claim is not simply higher-quality video generation, but a video-call-like experience: a virtual character can keep talking, moving, turning, reacting to the user’s voice and changing its future behavior as new input arrives.

Key points

• A unified streaming architecture: A1 places images, video, audio, historical frames and interaction history into a persistent state. This is meant to keep character identity, scene layout and object relationships stable during continuous generation.
• Interaction beyond transcribed text: Instead of relying only on ASR text, A1 models language meaning, acoustic features, non-verbal sound, environmental events, gestures and dynamic visual cues. Tone of voice or a sudden sound can influence content that has not yet been generated.
• Fast and slow control loops: According to the source material, the model can infer new response tokens within a minimal time slice of about 300 ms, while a higher-level Director Agent handles reasoning, tool use and longer-term behavior planning asynchronously.
• MJD for fewer diffusion steps: Vivix proposes Multidimensional Joint Distillation, compressing an 8-step quality baseline into a 2-step inference model while trying to preserve visual quality, motion range and long-horizon consistency.
• VMI inference infrastructure: The company describes an infrastructure stack that separates Encoder and Decoder services, introduces prefill/decode separation, redesigns KV Cache transfer, aligns training and inference for NVFP4, and co-schedules computation, communication and memory. Vivix reports over 10,000 video tokens/s on a single card, a shortest new-input response time of around 300 ms, and an average visible reaction latency below 0.6 seconds.

Why it matters

A1 points to a shift from lip-synced digital humans to real-time virtual characters with a sense of body. Such systems must handle identity persistence, motion generation, speech and event understanding, interruption, redirection and low-latency rendering at the same time. That is a harder problem than producing a short video clip offline.

If the reported capabilities hold up in broader use, the technology could affect interactive entertainment, AI companions, customer service avatars, game NPCs and real-time content creation. The benchmark for video AI may also change: not only whether a frame looks good, but whether the character can keep moving, respond instantly and avoid drifting over time.

Still, the available information comes mainly from company materials and media reporting. Independent tests, public demos and long-session evaluations will be needed to judge robustness. The direction is clear, but a reliable “parallel world” remains a long-term goal.

Source: QbitAI