Allegro

Allegro is an open-source text-to-video diffusion transformer from Rhymes AI, released under the Apache 2.0 license. It combines a 175M-parameter VideoVAE with a 2.8B-parameter VideoDiT to generate 6-second videos at 720x1280 and 15 FPS (interpolatable to 30 FPS with EMA-VFI). The newer Allegro-TI2V variant adds text-and-image-to-video, including subsequent-frame continuation and intermediate-frame generation between two existing shots — useful for extending or bridging clips. It ships with Diffusers pipeline support, multiple precision modes (FP32/BF16/FP16), and can run in BF16 with CPU offloading on as little as 9.3GB of GPU memory.

The honest picture: quality is solid for an open 3B-class model but visibly behind closed leaders like Veo and Kling on fine detail and complex motion. Generation is slow without serious hardware — about 20 minutes for a 6-second clip on a single H100, or 3 minutes on eight. It cannot render celebrities, legible text, or specific real-world locations. The ecosystem is research-grade: expect to wrangle Python environments and CUDA versions rather than clicking a web UI.

Clone github.com/rhymes-ai/Allegro or pip install the Diffusers AllegroPipeline. Requires Python 3.10+, PyTorch 2.4+, CUDA 12.4+. Checkpoints on Hugging Face (rhymes-ai/Allegro).