Vivix is an AI creation platform focused on real-time interactive multimodal models, offering Vivix-A1 (real-time full-duplex, generating full-body AI characters) and Vivix-W1 (interactive storytelling engine). Through MJD distillation and NVFP4 quantization, near-30B-parameter models run real-time inference on consumer GPUs with end-to-end latency below 1.7 seconds, and the API is now available.
What Is Vivix
Vivix is an AI creation platform focused on real-time interactive multimodal models. Its core products include: Vivix-A1, the world's first real-time full-duplex model, capable of generating interactive AI characters with full-body motion that can hear, speak, and act; Vivix-W1, a streaming native multimodal model that serves as an interactive storytelling engine enabling real-time plot changes through voice, text, and touch, along with the real-time video runtime Vivix IVI. Vivix's proprietary MJD distillation, NVFP4 quantization, and VMI inference engine enable near-30B-parameter models to run in real time on consumer-grade GPUs with interaction latency below 1 second. Vivix-A1's API is now available.
Key Features of Vivix
- Full-body character generation: Vivix-A1 can generate full-body AI digital humans in real time that can walk, turn, pick up objects, and dance.
- Full-duplex real-time interaction: Users can interrupt at any time by voice; the character's motion, expressions, and responses change instantly without regeneration.
- Synchronized audio-visual output: Speech, expressions, gestures, and ambient sound are natively unified in a single model with no stitching artifacts.
- Interactive storytelling generation: Vivix-W1 can generate an evolving audio-video story from a single sentence, and users can change the plot direction in real time.
- Multimodal input: Supports voice, text, images, clicks, touch, and other means of interaction.
- Live commerce applications: Digital hosts can move freely and showcase products, while operators can adjust the broadcast direction by voice in real time.
- Ultra-low latency: First-frame latency under 0.6 seconds, interruption response 300–900 milliseconds, end-to-end latency under 1.7 seconds.
- Real-time on consumer hardware: Near-30B activated-parameter models, via MJD distillation and NVFP4 quantization, can run real-time inference on a single consumer-grade GPU at a cost below $1 per hour.
How to Use Vivix
- Register and get an API Key: Visit the Vivix API platform at platform.vivix.ai, sign up for an account, and create an API Key in the developer console. All requests use
https://api.vivix.ai/as the base URL with Bearer authentication. - Choose a model: Select the A series (real-time Avatar interaction or script-driven Avatar video) or the W series as needed, such as
vivix-a1-stream,vivix-w1-stream, etc. - Create a session: Via the REST interface POST
/v1/realtime-avatar/sessions, configure parameters such as mode, character image, voice, persona instructions, and speech recognition (e.g., server_vad). - Establish a real-time connection: After obtaining the returned
session_id, control the session through the WebSocket control channel (wss://api.vivix.ai/v1/realtime-avatar/control); audio-visual output is streamed via RTC (TRTC). - Drive interaction: Write the user's voice or text into the conversation context; the Avatar generates replies in real time and streams them out as speech, video, and word-by-word subtitles, with support for mid-stream interruption.
- Real-time intervention: Through the Streaming Control interface, you can switch images/characters, send action commands, or directly script audio during generation without interruption.
- W1 world gameplay: Plot is driven by open Events; events can be cancelled before taking effect. Sessions support hold/pause and resume, allowing new events to be inserted at any time to alter the story direction.
Core Advantages of Vivix
- True full-duplex interaction: Breaking through the "half-duplex" request-response pattern, the model can catch new input at any time and instantly change content while continuously generating audio and video.
- Full-body character agency: A1 is the first real-time model to support complete full-body motion generation—characters can walk, turn, and pick up objects, rather than merely lip-syncing in a fixed shot.
- Native multimodal architecture: Speech, text, and images are modeled in a unified signal space, bypassing the ASR+LLM+TTS cascade, so perception directly drives behavior with naturally synchronized audio-visual motion.
- Ultra-low latency: First-frame latency < 0.6 seconds, interruption response 300–900 milliseconds, end-to-end < 1.7 seconds, overall interaction latency below 1 second.
- Large parameters + consumer-grade real-time: Near-30B activated parameters with an 8x8x4 high compression ratio, via MJD distillation and training-time NVFP4 quantization, achieve over 10,000 video tokens/s on a single consumer-grade GPU, with inference cost below $1 per hour.
- Natively interruptible streaming architecture: Causal generation + a VMI codec decoupling design lets new commands be appended directly in the decoding chain, so current generation need not be interrupted and restarted.
Vivix vs. Competitors
| Comparison | Vivix | HeyGen LiveAvatar |
|---|---|---|
| Technical approach | Native multimodal generative model (near-30B activated parameters); perception directly drives generation, not an ASR+LLM+TTS pipeline | Stitched pipeline: can plug in its own LLM/TTS (Lite mode) or use the full suite (Full mode); underlying layer is pre-rendered Avatar driving |
| Character expressiveness | Full-body motion generation: walking, turning, picking up objects, dancing, autonomous action in open worlds | Primarily upper-body presenter style; Avatar IV has full-body motion capture and micro-expressions but is limited to fixed studio scenarios |
| Narrative capability | W1 supports real-time interactive film/TV: generate a story from one sentence, users change the plot at any time | No such capability; single-character dialogue only |
| First-frame latency | TTFF < 0.6 s (official) | Median first frame < 300 ms |
| Interruption capability | Native interruption 300–900 ms; motion and response change instantly | Supports interruption handling, but manifests as "switch to listening animation + cancel current speech" |
| Real-time cost | Below $1/hour (official testing) | About $0.20/minute (about $12/hour, Full tier); at scale can drop to $0.01/minute |
| Language support | Specific language count not disclosed | 175+ languages with lip-sync; mature digital clone |
| Integration method | API + WebSocket control + RTC stream (A1 live, invite-only beta) | Pure API, requires building your own conversation stack; Credits expire monthly |
Application Scenarios of Vivix
- Live commerce: Full-body digital hosts appear and walk around, try on and pick up products, while operators adjust scripts and selections by voice in real time without interrupting the stream.
- Virtual companions / AI social: AI girlfriend/boyfriend apps where motion, expressions, and voice are naturally synchronized; users can interrupt conversations at any time and the character responds instantly.
- Interactive film/TV / interactive shorts: Generate a real-time evolving story from one sentence; viewers participate in plot decisions via voice and clicks, enabling a "everyone is a director" viewing mode.
- Game NPCs and open worlds: Endow game characters with autonomous perception and action; players interact naturally with NPCs via voice, advancing open-world narrative.
- Smart customer service and virtual shopping guides: 7×24 online full-body digital employees who can demonstrate products, answer questions, and respond to customer needs in real time, embeddable in websites/Apps/physical big screens.
Frequently Asked Questions about Vivix
Q1: What are the core models of Vivix? A1: Vivix-A1 is the world's first real-time full-duplex model, capable of generating interactive AI characters with full-body motion that can hear, speak, and act; Vivix-W1 is a streaming native multimodal model that provides an interactive storytelling engine and the real-time video runtime Vivix IVI, enabling real-time plot changes through voice, text, and touch.
Q2: What is the interaction latency of Vivix? A2: First-frame latency under 0.6 seconds, interruption response 300–900 milliseconds, end-to-end latency under 1.7 seconds, with overall interaction latency below 1 second.
Q3: How do I integrate Vivix? A3: Visit the Vivix API platform at platform.vivix.ai to register and create an API Key, create a real-time Avatar session via the REST interface, then drive audio-visual interaction through the WebSocket control channel and RTC stream.
Q4: Can consumer-grade GPUs run it in real time? A4: Yes. Near-30B activated-parameter models, via MJD distillation and NVFP4 quantization, can run real-time single-card inference on consumer-grade GPUs, with inference cost below $1 per hour.
Frequently asked questions
What are the core models of Vivix?
Vivix-A1 is the world's first real-time full-duplex model, capable of generating interactive AI characters with full-body motion that can hear, speak, and act; Vivix-W1 is a streaming native multimodal model that provides an interactive storytelling engine and the real-time video runtime Vivix IVI, enabling real-time plot changes through voice, text, and touch.
What is the interaction latency of Vivix?
First-frame latency under 0.6 seconds, interruption response 300–900 milliseconds, end-to-end latency under 1.7 seconds, with overall interaction latency below 1 second.
How do I integrate Vivix?
Visit the Vivix API platform at platform.vivix.ai to register and create an API Key, create a real-time Avatar session via the REST interface, then drive audio-visual interaction through the WebSocket control channel and RTC stream.
Can consumer-grade GPUs run it in real time?
Yes. Near-30B activated-parameter models, via MJD distillation and NVFP4 quantization, can run real-time single-card inference on consumer-grade GPUs, with inference cost below $1 per hour.
