GRPO post-trained Qwen2.5-Omni-3B variant used for small-model ablations in the AVSQwen release (run_abla/*_3B.sh). Same training recipe as the 7B flagship, trained 4200 steps.

Base model: Qwen/Qwen2.5-Omni-3B
Parameters: 3B
Training checkpoint: checkpoint-4200
Paper / code: https://github.com//AVSQwen

Usage

python

from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor

model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
    "Vegetabot/AVSQwen-Omni-3B",
    torch_dtype="auto",
    device_map="auto",
)
processor = Qwen2_5OmniProcessor.from_pretrained("Vegetabot/AVSQwen-Omni-3B")

For the full inference pipeline (frame selector + grounding + SAM2 segmenter), please refer to inference/ and run/*.sh in the release repo.

Training

Framework: ms-swift with GRPO
Rewards: bbox_format_reward, sam_keyframe_reward, minimal_efficiency_reward
Data: swift_train_bbox_grpo_balanced_full_v4_sam.jsonl (OmniAVS + 7 referring datasets, balanced mix)
See training/train_qwen25omni_3B_full_grpo_FINAL_V2.sh in the release repo.

Citation

bibtex

@article{avsqwen2026,
  title   = {AVSQwen: ...},
  author  = {...},
  year    = {2026}
}