---
license: cc-by-nc-4.0
library_name: libreyolo
pipeline_tag: any-to-any
base_model: sensenova/SenseNova-Vision-7B-MoT
tags:
  - libreyolo
  - unified-multimodal
  - object-detection
  - keypoint-detection
  - depth-estimation
  - image-segmentation
  - ocr
  - arxiv:2607.06560
---

# SenseNovaVision7b

> **Non-commercial weights.** These model weights are released by SenseTime
> under **CC BY-NC 4.0** and may be used for non-commercial purposes only.
> The LibreYOLO integration code is MIT; the vendored architecture code is
> Apache-2.0. The license of the weights does not change by being mirrored
> here.

SenseNova-Vision-7B-MoT, mirrored for LibreYOLO's `LibreVLM` tier. A unified
multimodal model (Bagel-MoT architecture: Qwen2.5-7B MoT decoder, SigLIP
vision tower, FLUX autoencoder) that serves many vision tasks from one
checkpoint: symbolic outputs (boxes, points, keypoints, OCR words) are
generated as tagged text, dense outputs (depth maps, segmentation masks,
panoptic maps) are generated as images decoded by the VAE.

## Source

Byte-identical mirror of
[sensenova/SenseNova-Vision-7B-MoT](https://huggingface.co/sensenova/SenseNova-Vision-7B-MoT)
at revision `79548fcc5b954598799b9317f8d3ec5e347d5c0e`.
Copyright (c) 2026 SenseTime Group Inc. and/or its affiliates.

Reference implementation:
[OpenSenseNova/SenseNova-Vision](https://github.com/OpenSenseNova/SenseNova-Vision)
(commit `12ccd96e32b32967a11cacb6c5bd5fe3a555fc0c`, Apache-2.0).
Paper: [Vision as Unified Multimodal Generation](https://arxiv.org/abs/2607.06560).

Weight checksums (SHA-256):

```
ema.safetensors  96f29abd98791288c5a24087322e964bb9bcabfc2f185ece71543f827bc2b11e
ae.safetensors   afc8e28272cd15db3919bacdb6918ce9c1ed22e96cb12c4d5ed0fba823529e38
```

## Modifications

None. The weights and configuration files are unmodified upstream bytes;
only this card and the NOTICE file are added.

One upstream quirk to be aware of: `tokenizer.json` assigns the chat/vision
special tokens ids beyond the checkpoint's embedding table, while
`tokenizer_config.json`'s `added_tokens_decoder` records the layout the model
was trained with (structured tokens overriding ids 149632-151664). LibreYOLO
reconstructs the trained layout at load time; other consumers should do the
same or use a legacy slow tokenizer built from `vocab.json`/`merges.txt`.

## Usage

```python
from libreyolo import LibreVLM

model = LibreVLM("sensenova-vision", task="detect")
model.set_classes(["person", "bicycle"])
results = model.predict("image.jpg")          # Results.boxes

model.set_task("depth")
results = model.predict("image.jpg")          # Results.depth_map

model.set_task("segment").set_classes(["the person on the left"])
results = model.predict("image.jpg")          # Results.masks
```

Supported LibreYOLO tasks: `detect`, `point`, `pose`, `ocr`, `depth`,
`segment` (referring), `panoptic`. Free-form access via `model.chat(...)`
and `model.generate(...)`.

## License Composition

- Model weights: **CC BY-NC 4.0** (non-commercial), SenseTime
- Reference implementation: Apache-2.0 (SenseTime, building on ByteDance's
  Bagel, Hugging Face transformers, and the Black Forest Labs FLUX
  autoencoder, all Apache-2.0)
- LibreYOLO integration: MIT

See [`LICENSE`](./LICENSE) (upstream license statement, verbatim) and
[`NOTICE`](./NOTICE) for attribution details.