What the repository says you can run on one card
The smallest published model in this register is a 5B text-to-video and image-to-video model at 720P and 24fps. The repository states it runs on consumer-grade cards such as a 4090, producing a five-second 720P clip in under nine minutes on a single GPU. As of 2026-09-12.
| Item | What is published | Detail |
|---|---|---|
| Model | 5B, text and image to video | 720P at 24fps |
| Card named | A consumer 4090 | The vendor's own example |
| Stated time | Under nine minutes | Five seconds of 720P output |
Inclusion rule. Figures stated in the vendor's own repository. Community benchmarks are not recorded. Order. As the repository presents them.
1Nine minutes for five seconds is the ratio that matters
Read as a ratio that is roughly a hundred seconds of compute for one second of output. Against a per-second price it is the number that decides whether local generation is cheaper, and it is one of just three timing figures published anywhere in this register.
It is also the vendor's own figure on the vendor's chosen command, which makes it a friendly case rather than an average. A planner should treat it as the shape of the answer rather than as their answer.
2A named card is more useful than a VRAM number
Saying a 4090 tells a reader immediately whether they have one or can rent one. A memory figure alone leaves them estimating, and estimating badly is how a plan ends up on hardware that swaps.
The same repository also states a much larger requirement for other commands, which is why this register keeps both figures rather than reporting whichever is more convenient.
3What running locally replaces and what it does not
It replaces a per-second rate with hardware, electricity and somebody's time. For a continuous workload that can be cheaper; for bursts it usually is not, because the card is idle between them and the rate is not.
It does not replace the version being sold. The published model here is a generation behind what the hosted tiers run, so this is a different product at a different cost rather than the same product cheaper.
- HardwareThe repository states TI2V-5B runs on consumer-grade graphics cards such as a 4090, generating a 5-second 720P video in under 9 minutes on a single consumer GPUvendor's own figure
- Smallest published modelA 5B text-to-video and image-to-video model at 720P and 24fpsTI2V-5B
4Open weights, hosted
Instead of local hardware for “Running it locally”, SceneMixer runs hosted models; its guides describe the result:
AI dubbing replaces the audio of a finished clip and then tries to match the lips. Native-language generation gives the video model the line in the target language before the shot exists, so mouth movement, pauses and shot length are produced for that language.
SceneMixer, Native-Language Dialogue vs. Dubbing in AI Video (2026): Why the Cast Should Speak the Line — “Is native-language generation the same as AI dubbing?” (read 2026-10-13)
The longest an episode may run: 240 seconds for Short Drama, 300 for TV Series, none for Film.
SceneMixer, Short Drama, TV Series or Film — “What is the episode cap?” (read 2026-10-13)
5Sources
Repository statements and hosted rates are linked from the model pages and the method page, checked 2026-09-12. Hardware figures are set out on requirements. Nearby: 4090 or 80GB, Clip length, Buy or run.