Vivix.AI builds real-time, interactive multimodal foundation models aimed at developers and creators who want AI characters that can be steered live rather than rendered as fixed clips. This launch introduces the A1 Playground and Agent, which lets anyone create a character that talks, moves, and plays with the user in a continuous session, with input coming through voice, text, or reference images. The underlying Vivix-A1 model moves beyond talking-head avatars to interactive AI characters that perceive and respond to their environment in real time, generating streaming multimodal output incrementally so characters can listen, reason, speak, and act continuously without cutting to separate clips.
The technical pitch for operators and infra buyers is full-duplex streaming video generation. Full duplex means it listens, responds and acts at the same time, and emotion shifts happen in one uncut take with no clip stitching, where most video models generate in 5 to 10 second blocks and drift over long runs, while A1 generates small incremental segments at 2 inference steps. That matters now because video generation has largely been a batch, prompt-to-clip workflow, and shipping low latency enough for a user to interrupt and redirect a character opens the door to game NPCs, companion apps, and live entertainment formats that pure text-to-video pipelines cannot support.
The company was founded in January 2025 by Liu Yu, the former executive research director at SenseTime , and is based in Hong Kong . On July 24, 2026, Vivix released two real-time interactive multimodal models, A1 and W1, each with approximately 30 billion activated parameters , and the A1 Playground and Agent release puts that research into the hands of end users and developers who want to build on top of the character layer rather than train their own.
Comments (7)
Sign in to join the discussion.
N
Nneka O.9d ago
The tweet copy is doing a lot of work but 'Live' is carrying the whole suitcase. If latency is above 400ms this whole pitch collapses into a chatbot with a face.
M
Marko Jelic9d ago
Nothing about pricing on the landing page, which usually means the answer is 'you can't afford it per minute.' GPU bills for real-time avatars are a special kind of character building.
T
tomas9d ago
Cute reveal video but every cut hides a loading state. Show me one unbroken 90 second session or I'm assuming the demo was lovingly hand-fed.
A
Ayesha M.9d ago
Watched this and immediately thought about moderation. 'Describe who you have in mind' is a sentence that will keep a trust and safety team employed for years.
K
Kenji Watanabe9d ago
Building in roughly this space and genuinely impressed by the voice sync in the clips. Curious if you're training your own base model or stitching on top of someone else's.
D
Deepak R.9d ago
A character you create is a toy. A character that remembers you across sessions is a product. Which one did you ship?
L
Lars V.9d ago
Onboarding better be 'type one sentence, see a face talking back' or Gen Z bounces in eleven seconds. The current signup flow has one too many fields for a product that sells vibes.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Vivix.AI
Building real-time interactive video models and the infrastructure to power them.
Building real-time interactive video models and the infrastructure to power them.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Vivix.AI builds real-time, interactive multimodal foundation models aimed at developers and creators who want AI characters that can be steered live rather than rendered as fixed clips. This launch introduces the A1 Playground and Agent, which lets anyone create a character that talks, moves, and plays with the user in a continuous session, with input coming through voice, text, or reference images. The underlying Vivix-A1 model moves beyond talking-head avatars to interactive AI characters that perceive and respond to their environment in real time, generating streaming multimodal output incrementally so characters can listen, reason, speak, and act continuously without cutting to separate clips. The technical pitch for operators and infra buyers is full-duplex streaming video generation. Full duplex means it listens, responds and acts at the same time, and emotion shifts happen in one uncut take with no clip stitching, where most video models generate in 5 to 10 second blocks and drift over long runs, while A1 generates small incremental segments at 2 inference steps. That matters now because video generation has largely been a batch, prompt-to-clip workflow, and shipping low latency enough for a user to interrupt and redirect a character opens the door to game NPCs, companion apps, and live entertainment formats that pure text-to-video pipelines cannot support. The company was founded in January 2025 by Liu Yu, the former executive research director at SenseTime , and is based in Hong Kong . On July 24, 2026, Vivix released two real-time interactive multimodal models, A1 and W1, each with approximately 30 billion activated parameters , and the A1 Playground and Agent release puts that research into the hands of end users and developers who want to build on top of the character layer rather than train their own.
Comments (7)
The tweet copy is doing a lot of work but 'Live' is carrying the whole suitcase. If latency is above 400ms this whole pitch collapses into a chatbot with a face.
Nothing about pricing on the landing page, which usually means the answer is 'you can't afford it per minute.' GPU bills for real-time avatars are a special kind of character building.
Cute reveal video but every cut hides a loading state. Show me one unbroken 90 second session or I'm assuming the demo was lovingly hand-fed.
Watched this and immediately thought about moderation. 'Describe who you have in mind' is a sentence that will keep a trust and safety team employed for years.
Building in roughly this space and genuinely impressed by the voice sync in the clips. Curious if you're training your own base model or stitching on top of someone else's.
A character you create is a toy. A character that remembers you across sessions is a product. Which one did you ship?
Onboarding better be 'type one sentence, see a face talking back' or Gen Z bounces in eleven seconds. The current signup flow has one too many fields for a product that sells vibes.