Visko Deploys Orbis Live Model on Google Cloud TPUs
The live inference pipeline now achieves 1.07x playback speed by utilizing Google Cloud TPU slices.
Updated on Oct. 28, 2026 in Artificial Intelligence
Live Poll
Do you believe real-time AI video interaction will improve your daily digital experience?
Visko has successfully deployed its Orbis live world model on Google Cloud TPUs, reaching a playback speed of 1.07x. The system generates continuous live video streams from text or image prompts using a specialized JAX, torchax, and Pallas software stack.
Why it matters
Meeting real-time playback deadlines required significant optimization, including rewriting the audio model in native JAX to achieve a 5.8x speed-up. The team also opted for BF16 precision over FP8 to maintain high video quality across continuous chunks.
The inference pipeline processes video and audio chunks sequentially on a single Cloud TPU slice using Ulysses context parallelism for token sequences. Worker restart times are kept under one minute by caching XLA compilation artefacts.
The players
Visko
Visko is an artificial intelligence research organization focused on developing advanced live world models.
Google Cloud
Google Cloud is a suite of cloud computing services that provides the infrastructure and TPU hardware for high-performance AI model training and inference.
The details
Engineers optimized the system by splitting token sequences across chips and computing attention in tiles. This configuration allows the three-model stack to share hardware efficiently for each playback cycle.
Timeline
The development phase from initial work to production handover spanned 30 days.
The Tech Race
The deployment follows a pattern established by the Google Cloud TPU accelerator architecture for handling high-throughput, low-latency AI inference workloads. This milestone reflects a broader industry shift toward optimizing native JAX environments to replace legacy frameworks.
Users can expect more stable and consistent high-quality live video streaming from prompt-based models as these optimization techniques become standard. The reduction in restart times means fewer interruptions for developers and consumers relying on the platform for live outputs.
The takeaway
Achieving real-time performance in generative world models depends heavily on granular compute budgeting and efficient hardware scheduling. Developers looking to optimize similar pipelines should prioritize native framework implementations to bypass legacy constraints.
Further reading
For more developments in generative systems, see the Artificial Intelligence section.
Source note: This article includes information reported by IT Brief Australia.
Live Poll
Do you believe real-time AI video interaction will improve your daily digital experience?
