Vimoo opens weights for an 11.7B-parameter video generator: 5 seconds at 720p

The open video generator market has gained another player, and this time the interesting part is less the model itself than how it was assembled. The company Vimoo has released the weights of the Vimoo 1.0 family for text-to-video (T2V) and image-to-video (I2V) generation. It did not train everything from scratch. Instead, it built a hybrid from ready-made open blocks around its own core.

What is inside

The core is a dual-stream diffusion transformer (Dual-Stream MMDiT) with 11.7 billion parameters. "Dual-stream" means that text and visual tokens travel along separate branches and interact through attention. Many modern image generators use a similar scheme.

Around the core sit third-party components:

• The video VAE from Wan 2.2 with 4×16×16 compression and 48 latent channels. It turns frames into a compact latent representation that the transformer works with.
• Qwen2.5-VL-7B handles the meaning of the prompt: what is in the scene and who is doing what.
• Glyph-ByT5-v2 is a character-level encoder for typography. Its job is to keep a sign or a headline in the frame from turning into a mush of pseudo-letters.
• Google SigLIP in I2V mode preserves the details of the reference frame, so the hero does not change on the third second.

Base generation yields 5 seconds: 121 frames at 24 FPS in 720p resolution. For 1080p there is a separate generative restoration model, Vimoo Enhancer. It does not simply stretch the picture but paints in additional details.

What it looks like in practice

Imagine you are making a promo clip for a coffee shop. You need a scene where a barista sets a cup on the counter and the name of the venue is legible on the glass behind. With a typical open generator, the text on the glass will almost certainly warp. Here a dedicated encoder is responsible for it, so the model has a chance to do better. How much better is still unclear, because the post contains no independent comparisons.

In I2V mode you provide a frame with a branded mug, and SigLIP helps preserve its look as it moves. Then the 720p draft can be run through the Enhancer to get 1080p.

The main benefit of this approach is that each block can be replaced or updated separately. A better encoder comes out, and you swap it in without retraining everything from scratch. For teams building their own video pipelines, this is a good architecture template.

Limitations and what to do about them

First, what we do not know. The post has no benchmarks, no generation examples and no comparison with Wan 2.2, on whose VAE the model is partly built. There is no data on video memory requirements or inference speed. The license for the weights also needs to be checked separately before bringing the model into a commercial product. Five seconds of video is a short clip, not a scene. For longer videos you will have to stitch fragments together and keep characters consistent.

What a developer or product manager should do:

• Download the weights from Hugging Face and run your own 10–20 typical prompts, especially ones with text in the frame. This is the only honest way to verify the claims.
• Compare the result with Wan 2.2 on the same requests. If the difference is small, there is no point in changing your stack.
• Read the technical report and the license before any production plans.
• If you are building your own generator, look at modular assembly as a way to save months of training.

The takeaway is simple. This is not a breakthrough but a careful engineering assembly of good open parts. It is interesting as a working example and as one more option for self-hosting. Whether Vimoo 1.0 becomes something bigger will only be shown by tests on real tasks.