HelixWorld needs ~70 GB of VRAM and has no hosted API: what to use

HelixWorld is a research model, not a service: you can run its open code only on your own Linux machine with roughly 70 GB of GPU memory, and its repository documents no hosted API. If you need finished video clips rather than a live interactive scene, you submit a generation job to a video model that is offered as a service, such as the models in Sume's catalog, and collect the file when the job completes. Sume does not list HelixWorld, so this post covers what it is, what it costs you to run, and where the line sits.

What is HelixWorld?

The HelixWorld repository describes it as a real-time interactive audio-visual world model: the visual scene and camera-grounded spatial stereo sound change together as you move through an environment. The technical report, submitted on 2026-09-29, says it distills a bidirectional teacher into a faster streaming student and takes 6-DoF camera trajectories and user actions as conditioning.

The repository's release dates are 2026-09-03 for the Preview v1 code and checkpoint, 2026-09-29 for the report, and 2026-10-12 for HelixBench, a benchmark of whether synthesized sound fields track camera motion.

What does it take to run it yourself?

Should you use it for a marketing clip?

Probably not today. HelixWorld is built for interaction: you steer a camera and the scene responds. A marketing clip is a fixed file with a prompt, a duration and an aspect ratio. The report's headline claim is that it matches silent world models on visual fidelity and beats baselines on camera-aligned stereo sound, which matters if spatial audio is your deliverable and is a poor reason to rent an 80 GB GPU if it is not.

Treat the numbers as the authors' own. We have not run HelixBench, and nothing here is an independent measurement.

How does the job route differ?

On Sume, a generation request creates a durable job and returns its id at once. You poll with backoff until the job is completed, failed or canceled, and you do not resubmit a paid request if your own process times out, because the id stays valid. A webhook can deliver the terminal event instead of polling. Results can be read only after completion, and several jobs can be read in one batch call. The details are on the Jobs and results page.

For parallel clips see 60 parallel clips in batches of 20. For another preview model with the same question, see Reka Rho-1.

When would self-hosting HelixWorld make sense?

Pick it when you are doing research, need a live explorable scene with synchronized stereo sound, and already have an 80 GB-class GPU. Pick a hosted job route when you need a predictable file per request and no GPU to babysit. If HelixWorld later ships a hosted endpoint, check its own page for pricing and limits before planning around it; none is published in the sources read today.