The fit problem comes first

Most open video models are open in the sense that the files exist. Vimoo 1.0's files exist at a size that excludes the hardware most local users own. The two transformers, the glyph encoder and the Qwen2.5-VL-7B text encoder add up to 64.3 GB, and WaveCut reports BF16 out-of-memory results on its tested RTX 5090 configuration, with and without model offload. On a 96 GB RTX PRO 6000 it runs, at about 94 minutes per default clip.

The Lite model has the same problem in miniature. Its original 36.8 GB also runs out of memory on the 5090 in WaveCut's testing. So for both sizes, on the configuration WaveCut tested, the quantized build is not an optimization. It is the way the model runs on that card at all. That does not rule out other implementations, configurations or quantization methods.

Then the clock problem

The 4-bit full model fits, with room to spare: 25.2 GiB peak, dropping to 17.8 GiB with offload. What it does not do is finish quickly. The listed text-to-video time is 49.7 minutes for five seconds of 720p video, and 31.1 minutes with INT8 attention turned on.

Read the measurement note before quoting either figure. For the full model, the card states that the configurations ran 6 timed steps after the warm-up instead of 50, because a complete BF16 comparison run takes about 94 minutes on the larger card. The tables extrapolate to 50 steps: the measured time plus 44 times the measured time per step (59.1 seconds per step on the 5090). The 49.7 minutes is that projection, not a stopwatch on a finished clip. It is clearly labeled arithmetic on the card, but it is still an estimate.

The Lite card carries no extrapolation note; its figures are WaveCut-reported timings for the same clip shape. A five-second clip in 4 minutes 36 seconds, or 3 minutes 21 seconds with INT8 attention, inside 20.5 GiB, is a substantially shorter turnaround. Whether either build is practical depends on workload, quality needs, hardware and operating constraints, but on the reported numbers Lite is the plausible local candidate in a way the flagship is not.

What the 4 bits cost

WaveCut evaluated the quantized builds against the BF16 originals on prompts kept out of the calibration set. WaveCut reports closer agreement with BF16 on its image-to-video comparisons than on text-to-video: LPIPS 0.094 on the full model and 0.113 on Lite, with PSNR near 35 dB, against LPIPS 0.368 to 0.377 and PSNR under 18 dB for text-to-video. These metrics measure similarity to the BF16 reference, not task quality or human preference. WaveCut's explanation is that over 50 sampling steps a small numerical difference changes the path of the motion, so the scene, subjects and camera move survive while the movement itself takes a different route, and a distance metric counts that as error even when a viewer might not.

One detail is worth keeping, within the scope WaveCut tested: swapping the 4-bit text encoder back to the original BF16 encoder did not close the gap, which points at the sampling path rather than the prompt encoding in that configuration.

The practical angle

Vimoo-AI also sells hosted access, including a dedicated e-commerce mode that builds 20-second vertical product videos from one to nine reference images, and its API documentation lists the Lite model as currently free. That sets the local calculation. Running Lite at home costs electricity and patience, avoids a per-clip bill, and can keep generation local, subject to the software and integrations chosen. Running the full model at home, at roughly ten minutes of compute per second of output on the card WaveCut tested, is a different proposition, and the free hosted Lite tier is a different deployment option rather than evidence against local use.

Sources

  1. [1] WaveCut, Vimoo-1.0-OrbitQuant-W4A8 model card, Hugging Face (primary source for sizes, memory, timings, quality and the extrapolation note)Read source
  2. [2] WaveCut, Vimoo-1.0-Lite-OrbitQuant-W4A8 model card, Hugging Face (primary source for Lite figures)Read source
  3. [3] WaveCut, Vimoo — Orbit Quant collection, Hugging Face (update timing, November 15, 2026)Read source
  4. [4] Vimoo API documentation (hosted tiers, e-commerce mode, Lite listed as free; checked November 15, 2026)Read source