attashe/Bernini-MLLM-Qwen2.5-VL-7B

Bernini MLLM (Qwen2.5-VL-7B)

This is the MLLM semantic planner from ByteDance Bernini (Bernini: Latent Semantic Planning for Video Diffusion), a fine-tuned Qwen2.5-VL-7B-Instruct, extracted into a standalone transformers-loadable checkpoint.

In the original release the planner weights are bundled inside the unified `ByteDance/Bernini-Diffusers` checkpoint under the mllm. prefix. This repo simply extracts those tensors (prefix stripped, cast to bfloat16) and pairs them with the standard Qwen2.5-VL config / tokenizer / processor, so the model loads with stock transformers like any other Qwen2.5-VL model. The weights are unchanged.

Usage

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "attashe/Bernini-MLLM-Qwen2.5-VL-7B", dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained("attashe/Bernini-MLLM-Qwen2.5-VL-7B")

Notes

• —Architecture: Qwen2_5_VLForConditionalGeneration (8.29B params), bfloat16.
• —These are ByteDance's fine-tuned Bernini planner weights; within the full Bernini pipeline the planner's hidden states feed a DiT renderer, so as a standalone chat/VL model its behaviour may differ from the base Qwen2.5-VL-7B-Instruct.
• —License: Apache-2.0, inherited from the upstream Bernini release.

Citation

@article{bernini,
  title   = {Bernini: Latent Semantic Planning for Video Diffusion},
  author  = {Chenchen Liu and Junyi Chen and Lei Li and Lu Chi and Mingzhen Sun and Zhuoying Li and Yi Fu and Ruoyu Guo and Yiheng Wu and Ge Bai and Zehuan Yuan},
  journal = {arXiv preprint arXiv:2605.22344},
  year    = {2026}
}