Allegro open-sources a smaller lab's video model

The shot

Rhymes AI, a smaller video-generation lab, open-sourced Allegro on 22 October 2024, followed by an image-conditioned variant, Allegro-TI2V, on 25 November 2024. The company's own technical report, posted to arXiv on 20 October 2024 under the title "Allegro: Open the Black Box of Commercial-Level Video Generation Model," describes a 2.8-billion-parameter diffusion transformer paired with a 175-million-parameter video autoencoder, generating video up to six seconds at 15 frames per second and 720p resolution. As retrieved on 16 September 2026, the repository shows the project has since added training code and lower-resolution research checkpoints.

What the documents show

The paper states that "our user study shows that Allegro surpasses existing open-source models and most commercial models, ranking just behind Hailuo and Kling." This is Rhymes AI's own user study, run and reported by the company that built the model, not an independently conducted evaluation, and the paper's abstract does not describe the study's participant pool or scoring method. The repository's license file is a standard Apache License, Version 2.0, which the README confirms covers the repository without a registration step or revenue-based restriction, distinguishing Allegro's terms from the tiered licensing seen on some larger open releases in this pack.

The workflow

Because Allegro carries an unrestricted Apache 2.0 grant, a production or tool-builder can redistribute, modify, and deploy it commercially without the registration or revenue-threshold steps that govern some peer open releases, provided ordinary attribution terms are met. What remains a human decision is quality verification: the vendor's ranking claim, placing Allegro just behind two named commercial systems, is not something a production can rely on without running its own comparison against the specific footage type it needs.

What the tool does not change

A self-reported user study, however favorably worded, is not the same evidentiary category as a third-party benchmark conducted under published, reproducible conditions; this site treats a developer's own preference study as a vendor claim requiring that label, not as confirmed general capability, whatever a permissive license allows a team to build with the model.

• What methodology did Rhymes AI's cited user study use, and is that methodology published in enough detail to assess?
• Does Allegro's stated maximum of six seconds at 720p meet a specific production's shot requirements?
• Has any party outside Rhymes AI independently replicated the paper's ranking against Hailuo and Kling?

Allegro's Apache 2.0 license removes a legal obstacle that some larger open releases retain, but it does not convert the developer's own comparison into independent proof of where the model actually ranks.