Allegro
Apache-licensed diffusion transformer that generates 6-second 720p videos from text prompts.
About
Six seconds of 720p video from a text prompt is what Allegro produces, an open text-to-video release from Rhymes AI built as a diffusion transformer with a 175M-parameter VAE and a 2.8B-parameter DiT. Output is 88 frames at 15 fps in 720x1280 resolution, and the Allegro-TI2V variant adds image-to-video generation conditioned on a first frame and an optional last frame, alongside smaller 40-frame research checkpoints at 720p and 360p. Generation is not fast: around 20 minutes for one clip on a single H100, dropping to roughly 3 minutes across eight H100s, and VRAM needs are 9.3 GB with CPU offloading enabled or 27.5 GB without. Setup requires Python 3.10 or newer, PyTorch 2.4 or newer, and CUDA 12.4 or newer, installed from the GitHub repository with weights on Hugging Face. Both code and weights are Apache-2.0, permitting commercial use, which distinguished it among the fully open video model releases of late 2024. The technical report is on arXiv (2410.15458), and a public gallery shows sample generations.
Should you use Allegro?
Pick it when
Pick Allegro when you need a text-to-video model whose code and weights are both Apache-2.0 for commercial work, have about 10 GB of VRAM with CPU offload, and can queue jobs offline, since one 6-second 720p clip takes about 20 minutes even on an H100.
Look elsewhere when
Skip it if speed matters: 15 fps output and roughly 20 minutes per clip on an H100 is slow next to distilled LTX-Video or FastVideo's FastWan models. For a permissive license with more modes, Wan 2.1 or Kandinsky 5.0 Lite is the stronger default.
Alternatives to Allegro
- Wan 2.1
Also Apache-2.0 but a broader suite, 1.3B to 14B with image-to-video, first-last-frame, and VACE editing, and the 1.3B model runs in about 8 GB.
- Kandinsky 5.0
MIT-licensed, and its distilled 2B Lite model renders a clip in about 35 seconds on an H100 versus Allegro's 20 minutes, plus a 1080p Pro tier.
- LTX-Video
Far faster, with distilled checkpoints needing as few as 8 steps plus keyframe and extension control, but its weights use OpenRail-M or Lightricks' open weights license, not Apache-2.0.
- CogVideoX
The 2B model is also Apache-2.0 and runs in about 4 GB with quantization, but it outputs 720x480 rather than Allegro's 720p.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Video Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- Apache-2.0
- Minimum VRAM
- 10 GB
- Added
- Jul 29, 2026