Vivix-W1

Unveiling Vivix-W1: Toward a Streaming-Native Multimodal Interactive Model for Narrative Intelligence

In text-based interaction, humans typically process only ~3–5 tokens per second. In contrast, video-based interaction delivers a continuous multimodal stream composed of tens of thousands of visual tokens, along with synchronized audio tokens, every second. This represents an approximately 10³–10⁴-fold increase in effective token throughput. As a result, real-time video systems must handle enormous information throughput while maintaining millisecond-level latency, continuously perceiving, generating, updating memory, and responding to user input within a unified loop.

Under this regime, the traditional architecture of video generation systems—centered around offline pipelines that follow a static “input prompt → inference → output clip” paradigm—is not well suited to continuous real-time interaction. Real-time interactive video demands a new class of models that are inherently streaming-native, capable of continuously interpreting evolving user intent while simultaneously generating coherent audiovisual content in an ongoing, closed-loop process.

Today, we introduce Vivix-W1, a streaming-native multimodal interactive model with multi-shot storytelling, audio-visual co-generation, and real-time interaction. It is capable of preserving long-horizon consistency across evolving scenes while maintaining fine-grained responsiveness at second-level latency.

Streaming Audio-Visual Omni-Model

Vivix-W1 departs from the standard clip-level diffusion rendering paradigm and natively combines the understanding of interaction signals with a streaming generation architecture.

Capability 2: Multimodal Reference Support

Vivix-W1 supports multimodal references, including images, videos, audio, text, and interaction context. This allows Vivix-W1 to generate worlds that are both controllable and dynamic.

Capability 3: Interactive Signal Understanding with Real-Time Streaming Generation

Vivix-W1 uses streaming generation. Instead of generating the entire video as a single closed clip, it generates the world continuously through fine-grained audiovisual segments. This fine-grained streaming approach allows Vivix-W1 to integrate diverse interaction signals directly into the generation process, so each audiovisual segment can be driven and adjusted in real time by user interaction.

Capability 4: Native Multi-Shot Narrative Generation

Vivix-W1 introduces native multi-shot narrative generation. Unlike single-shot video generation systems, Vivix-W1 is designed to understand and generate coherent multi-shot sequences natively. It can coordinate scene cuts, camera motion, dialogue rhythm, character reactions, and multi-character interactions, allowing the generated content to feel like a coherently directed sequence rather than a collection of isolated clips.

Capability 5: Long-Horizon Streaming Generation for Continuous Narratives

To enable stable long-horizon real-time generation, we introduce Vivix-Turbo, a dedicated acceleration and consistency optimization technology for Vivix-W1. Vivix-Turbo is designed to push streaming world generation toward highly efficient real-time inference while maintaining long-horizon stability, multimodal consistency, and narrative continuity.

Vivix-Turbo optimizes the system across four key dimensions. First, it applies ultra-low-step distillation to reduce inference cost and make real-time streaming practical. Second, it jointly optimizes shot planning and temporal continuity over long sequences, allowing the model to generate extended multi-shot sequences without breaking narrative flow. Third, it reduces error accumulation over long streaming horizons, improving identity stability, scene consistency, camera coherence, and motion continuity. Finally, it strengthens multimodal reference consistency, ensuring that characters, voices, styles, environments, and behavioral patterns remain stable as the generated world continues to evolve.

Together, these optimizations allow Vivix-W1 to generate continuous, interaction-ready audio-visual streams with greater stability over extended durations. Instead of producing isolated clips, Vivix-W1 can maintain a continuously evolving generated world that remains coherent, responsive, and visually expressive over time.

Based on this capability, we can efficiently construct the following long-sequence audio and video streams within a time budget of less than one minute.

Limitations and Future Work

Validated at a medium training scale with approximately 30B active parameters, Vivix-W1 demonstrates a broad set of capabilities including multimodal understanding, real-time interaction, native audio-visual generation, multi-shot storytelling, and continuous streaming.

That said, Vivix-W1 is only the beginning. In certain complex scenarios, especially those requiring professional-grade visual fidelity, complex motion, or dense scene dynamics, there remains a gap compared with state-of-the-art clip-level video generation models trained on datasets containing billions of examples. These models benefit from significantly larger professional datasets and longer optimization cycles, especially for offline high-quality video generation.

We are now scaling Vivix-W1’s data engine toward tens of millions of hours of training data, expanding the coverage of cinematic scenes, multi-shot narratives, audio-visual events, character actions, spatial dynamics, and interaction patterns. As the data scale continues to grow, we expect Vivix-W1 to deliver significantly stronger stability, visual quality, complex motion, narrative coherence, and real-time interaction performance.

With full-scale training already underway for future versions, these results offer a compelling preview of where omni-modal interactive video is headed.