Back to Gap Map Model components / Multimodal models

Ovis

ATH-MaaS
open weights / Overall score: 2.7

Ovis (Open VISion) is a multimodal model line, first published by Alibaba's AIDC-AI team and now under the ATH-MaaS account, built on an architecture that structurally aligns visual and textual embeddings. Ovis2.6, the current release, moves the language backbone to a mixture-of-experts model with about 3B active parameters, reads images up to 2880×2880 pixels in a 64K-token context, and can crop and rotate an image while it reasons about it.

OvisOCR, Ovis-U1, Ovis-Image and the Ovis embedding models share the name but are separate models and are not part of this entry.

Openness

3 high confidence
3.0
weights
open (Ovis2.6-30B-A3B and 80B-A3B on the Hub, ungated)
data
closed (no training data released for Ovis2.6)
code
partial (fine-tuning code and an SFT script for Ovis2.5; nothing specific to Ovis2.6)
license
Apache-2.0 (code and weights)

Ovis2.6's weights and the repository code are Apache-2.0. The training data is not released, and the repository's fine-tuning scripts target the earlier Ovis2.5 rather than the current release.

Adoption

2 high confidence
2.0

Hugging Face downloads over the trailing 30 days for the six declared checkpoints across Ovis2.6, Ovis2.5 and Ovis2. OvisOCR, a separate document model, is downloaded more than all of them and is not counted.

Capability

3 high confidence
3.0

Ovis2.6 reads documents, charts, several images and video, level with MiniCPM-V. Its image tools crop and rotate within its own reasoning and do not act in a live interface.

Verified 2026-11-12