File size: 2,459 Bytes
9219cb6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
---
license: other
license_name: verus-noncommercial-research-evaluation
license_link: LICENSE
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3-VL-8B-Instruct
datasets:
  - Verus-Team/Verus-Bench
tags:
  - verus
  - video-understanding
  - visual-artifact-detection
  - safetensors
---

# Verus-Insight-8B

Verus-Insight-8B is a vision-language model from **Verus-Team** for inspecting
visual artifacts in videos. It produces text responses from visual inputs and instructions.

- **Model:** [Verus-Team/Verus-Insight-8B](https://huggingface.co/Verus-Team/Verus-Insight-8B)
- **Code, prompts and environment:** [Verus-Team/Verus-Insight](https://github.com/Verus-Team/Verus-Insight)
- **Benchmark:** [Verus-Team/Verus-Bench](https://huggingface.co/datasets/Verus-Team/Verus-Bench)

## Model files

- Architecture: `Qwen3VLForConditionalGeneration`, based on Qwen3-VL-8B-Instruct.
- Weight precision: BF16.
- Format: four Safetensors shards with a weight index.
- Tokenizer, image and video processors, and chat templates are included.

## Loading

Install the environment described in the [code repository](https://github.com/Verus-Team/Verus-Insight).
Native inference has been verified with Transformers 5.12.1 and PyTorch 2.11.0
on an NVIDIA H100 80GB GPU. See the code repository for the separate vLLM environment.

```python
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "Verus-Team/Verus-Insight-8B"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    dtype="bfloat16",
    attn_implementation="sdpa",
).to("cuda:0").eval()
```

Weights occupy approximately 17.53 GB;
inference also needs memory for visual inputs, activations and the key-value cache.
The code repository provides inference instructions and task prompts. The chat
template in this repository defines message formatting.

## Limitations

Outputs can miss visual artifacts or report artifacts that are absent. Results
depend on prompts, frame sampling and image resolution. When using Verus-Bench,
follow its documented input ordering and timestamps.

## License

Verus-Insight is provided for **non-commercial research and evaluation only**
under the [Verus license](LICENSE). Upstream Qwen3-VL components retain their
[Apache License 2.0](LICENSE-APACHE-2.0). See [NOTICE](NOTICE) for attribution
and the scope of the Verus modifications.