Technology / WORLD

Visko streams a live world model in real time on one TPU slice

Visko and Google Cloud teams moved the full inference pipeline for Orbis, a continuously streaming video world model, onto one Cloud TPU slice. Video generation, super-resolution and synchronized audio run sequentially on the accelerator. The companies say the work took 30 days from kickoff to production handover.

Reading sizeSources ↓Share story ↗
Conceptual illustration of technology, processors and connected information.Illustration
Symbolic illustration

The source reports output at 1.07 times playback speed and 1664×936 resolution at 18 frames per second, with about 7% headroom for streaming variance. This is a single-slice test described by the company, not an independent benchmark. The stack uses JAX, torchax and custom Pallas kernels.

Data

Timeline

Visko streams a live world model in real time on one TPU slice

Visko and Google Cloud teams moved the full inference pipeline for Orbis, a continuously streaming video world model, onto one Cloud TPU slice. Video generation, super-resolution and synchronized audio run sequentially on the accelerator. The companies say the work took 30 days from kickoff to production handover. The source reports output at 1.07 times playback speed and 1664×936 resolution at 18 frames per second, with about 7% headroom for streaming variance. This is a single-slice test described by the company, not an independent benchmark. The stack uses JAX, torchax and custom Pallas kernels.

  1. · Occurred

    Visko streams a live world model in real time on one TPU slice

    Visko and Google Cloud teams moved the full inference pipeline for Orbis, a continuously streaming video world model, onto one Cloud TPU slice. Video generation, super-resolution and synchronized audio run sequentially on the accelerator. The companies say the work took 30 days from kickoff to production handover. The source reports output at 1.07 times playback speed and 1664×936 resolution at 18 frames per second, with about 7% headroom for streaming variance. This is a single-slice test described by the company, not an independent benchmark. The stack uses JAX, torchax and custom Pallas kernels.

Methodology

Reported values and benefits are attributed to the source; no independent measurement was made.

Updated:

Structured coverage

Subjects and places

Subject proposals

Taxonomy version: IPTC Media Topics 2026-Q2

Continue reading

Technology

Cloudflare releases Clef-omni for unified audio and video decisions

Cloudflare announced Clef-omni, an open-weight decision model that accepts audio and video alongside text and images. Instead of chaining separate speech transcription and video-frame models, it classifies and selects field values from multiple modalities within one request.

Cloudflare1 min read Data included