Aria
Rhymes AIAria is Rhymes AI's open multimodal mixture-of-experts model, pretrained natively on mixed text, image and video input. It has 25.3B total parameters with 3.9B active per token and a 64K-token multimodal context, and is aimed above all at video and document understanding. Base checkpoints with 8K and 64K context and a dialog-tuned Aria-Chat are released alongside it.
Openness
3 high confidence- weights
- open (Aria, Aria-Chat and the Aria-Base-8K and 64K checkpoints on the Hub, ungated)
- data
- closed (no pretraining or post-training data released)
- code
- partial (inference and fine-tuning code with LoRA and full-parameter recipes; no pretraining pipeline)
- license
- Apache-2.0 (code and weights)
Aria's code and weights are Apache-2.0. Rhymes AI has released fine-tuning code but neither its training data nor the pipeline it trained the model with.
- https://huggingface.co/api/models?author=rhymes-ai&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt&expand[]=lastModified recorded 2026-11-12
rhymes-ai/Aria, Aria-Chat, Aria-Base-8K and Aria-Base-64K all gated false and tagged apache-2.0.
- https://raw.githubusercontent.com/rhymes-ai/Aria/main/LICENSE recorded 2026-11-12
LICENSE is the Apache License, Version 2.0.
- https://raw.githubusercontent.com/rhymes-ai/Aria/main/README.md recorded 2026-11-12
README Fine-tuning section: "We offer both LoRA fine-tuning and full parameter tuning, using various dataset types"; the base models "are fully compatible with this inference & fine-tuning codebase"; no pretraining instructions.
- https://ungh.cc/repos/rhymes-ai/Aria/files/main recorded 2026-11-12
Repository tree: aria/train.py, recipes/config_full.yaml and config_lora.yaml, and per-dataset example configs; no pretraining code or data.
Adoption
2 high confidenceHugging Face downloads over the trailing 30 days for the four declared checkpoints, almost all of them for Aria itself.
- https://huggingface.co/api/models?author=rhymes-ai&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt&expand[]=lastModified recorded 2026-11-12
rhymes-ai/Aria 17,956, Aria-Base-8K 113, Aria-Base-64K 112 and Aria-Chat 104 downloads in the trailing 30 days (18,285 together).
Capability
3 high confidenceAria reads documents, charts, several images and long videos, level with MiniCPM-V. It does not act in a live interface or take audio input.
- https://huggingface.co/rhymes-ai/Aria/raw/main/README.md recorded 2026-11-12
Aria card: "superior in video and document understanding"; "Aria supports multimodal input of up to 64K tokens. It can caption a 256-frame video in 10 seconds."; MMMU 54.9, ChartQA 86.4 and LongVideoBench 65.3 for Aria.
- https://raw.githubusercontent.com/rhymes-ai/Aria/main/README.md recorded 2026-11-12
README Fine-tuning section lists multi-image datasets and video datasets among the supported dataset types.
Verified 2026-11-12