Verus-Insight-8B / README.md
SuperCarryZZR's picture
Release Verus-Insight-8B
9219cb6 verified
|
Raw History Blame Contribute Delete
2.46 kB
metadata
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 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. 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.

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. Upstream Qwen3-VL components retain their Apache License 2.0. See NOTICE for attribution and the scope of the Verus modifications.