CRISPR — Checkpoints

Checkpoints for CRISPR: Context-Refined Information Spatial Pooling with Region-awareness for Efficient Visual Token Compression in VLMs, accepted at ACM MM 2026.

• Code: https://github.com/ZuyiZhou/CRISPR
• Paper DOI: https://doi.org/10.1145/3767308.3835007

The original release contains CRISPR checkpoints for the Qwen2.5-VL-3B-Instruct backbone (9x and 16x compression) and the Qwen2.5-VL-7B-Instruct backbone (16x compression). The 7B/9x checkpoint is not currently available (lost prior to this release) and is not planned unless retraining happens in the future.

Qwen3.5 checkpoints

Additional CRISPR compressor checkpoints are available for the frozen Qwen3.5 backbones, with a fixed visual-token budget per image.

Backbone	Visual tokens per image	Checkpoint
Qwen3.5-4B	16	qwen3.5-4b-16tokens
Qwen3.5-4B	36	qwen3.5-4b-36tokens
Qwen3.5-4B	64	qwen3.5-4b-64tokens
Qwen3.5-9B	36	qwen3.5-9b-36tokens
Qwen3.5-9B	64	qwen3.5-9b-64tokens

Each directory contains model.safetensors (float32 compressor weights), config.json (inference settings), and SHA256SUMS. The frozen Qwen3.5 backbone is downloaded separately. These files contain only model tensors and inference metadata; optimizer state and training metadata are excluded.

For the loader and compressor interface, see the Qwen3.5 guide. The legacy create_model_v7 loader below applies to Qwen2.5-VL only. File sizes and checksums are listed in qwen35_manifest.json.

Qwen2.5-VL checkpoint files

Path	Compression ratio	Backbone	Notes
3b_9x/checkpoint.pt	9x (3x3 block)	Qwen2.5-VL-3B-Instruct	Stage-2, best checkpoint by val loss
3b_9x/config.json			training config used to produce this checkpoint
3b_16x/checkpoint.pt	16x (4x4 block)	Qwen2.5-VL-3B-Instruct	Stage-2, best checkpoint by val loss
3b_16x/config.json			training config used to produce this checkpoint
7b_16x/checkpoint.pt	16x (4x4 block)	Qwen2.5-VL-7B-Instruct	Stage-2, best checkpoint by val loss
7b_16x/config.json			training config used to produce this checkpoint

Each checkpoint.pt is a plain torch.save dict with keys config, token_mixer (TokenMixer state dict), and local_c3 (LocalC3 state dict, which also contains the Global Token Fusion sub-module). Only the trainable CRISPR modules are included — the frozen Qwen2.5-VL vision encoder and decoder weights are not part of this checkpoint and must be obtained separately from Qwen2.5-VL. Optimizer/ scheduler state is not included (only the model weights needed for inference or further fine-tuning are provided).

Qwen2.5-VL usage

from crispr import create_model_v7

model = create_model_v7(decoder_path="./Qwen/Qwen2.5-VL-3B-Instruct")
model.load_checkpoint("3b_9x/checkpoint.pt")  # see crispr/model_v7.py for the loader
# for the 7B backbone: decoder_path="./Qwen/Qwen2.5-VL-7B-Instruct", checkpoint="7b_16x/checkpoint.pt"

See the main repository (https://github.com/ZuyiZhou/CRISPR) for the model code, training script, and evaluation scripts.

Citation

@inproceedings{zhou2026crispr,
  author    = {Zhou, Zuyi and Xue, Dizhan and Qian, Shengsheng and Xu, Changsheng},
  title     = {CRISPR: Context-Refined Information Spatial Pooling with
               Region-awareness for Efficient Visual Token Compression in VLMs},
  booktitle = {Proceedings of the 34th ACM International Conference on
               Multimedia (MM '26)},
  year      = {2026},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  doi       = {10.1145/3767308.3835007}
}