CoTinyVLA-Qwen3.5-0.8B

CoTinyVLA

CoTinyVLA is a compact vision-language-action policy built on the Qwen3.5-0.8B backbone. The full policy has approximately 0.9B parameters.

Model summary

• Input: 8 third-person frames, 8 wrist-camera frames, 8-D proprioception, and a natural-language instruction
• Reasoning: one episode-level Plan and one chunk-level Think span
• Output: an 8-step action chunk with 7 action dimensions per step
• Training: two epochs on the union of the four LIBERO-Plus suites
• Final checkpoint step: 102,024

Results

LIBERO-Plus

Standard LIBERO

Files

• student_state_dict.pt: complete policy state dictionary
• action_head_config.json: policy and reasoning-token configuration
• tokenizer and processor files: multimodal preprocessing and added tokens
• release_info.json: release metadata
• SHA256SUMS: file-integrity checksums

Usage

This checkpoint uses a custom policy class and is not directly loadable with AutoModel.from_pretrained.

Download the checkpoint with:

hf download euphoria-64/CoTinyVLA-Qwen3.5-0.8B --local-dir checkpoints/cotinyvla

Training, inference, and evaluation code:

https://github.com/BrainJellyPie/CoTinyVLA

Limitations

• Evaluation is limited to LIBERO and LIBERO-Plus simulation.
• Sim-to-real and cross-embodiment transfer have not been evaluated.
• The current implementation uses synchronous reasoning generation.
• Physical-robot deployment requires additional safety validation.

License and third-party resources

This repository does not redistribute LIBERO or LIBERO-Plus datasets, assets, or environment code. Users must obtain them from their original sources and follow their respective licenses.

The final checkpoint license will be specified before public release.

<!-- PAPER_CITATION_START -->

Paper

CoTinyVLA: Chain-of-Thought Distillation for a Sub-Billion-Parameter Vision-Language-Action Model

Minhyeok Lee, Chiyoung Kim, Chanhoe Gu, Seongrok Kim, Sanghyuk Roy Choi, Donghwan Hwang, Donghun Ryu, and Seokhyun Kim.

• Paper: https://arxiv.org/abs/2607.25487
• DOI: https://doi.org/10.48550/arXiv.2607.25487
• Code: https://github.com/BrainJellyPie/CoTinyVLA

Citation

Please cite the following paper when using this checkpoint:

@article{lee2026cotinyvla,
  title   = {CoTinyVLA: Chain-of-Thought Distillation for a Sub-Billion-Parameter Vision-Language-Action Model},
  author  = {Lee, Minhyeok and Kim, Chiyoung and Gu, Chanhoe and Kim, Seongrok and Choi, Sanghyuk Roy and Hwang, Donghwan and Ryu, Donghun and Kim, Seokhyun},
  journal = {arXiv preprint arXiv:2607.25487},
  year    = {2026},
  doi     = {10.48550/arXiv.2607.25487},
  url     = {https://arxiv.org/abs/2607.25487}
}

<!-- PAPER_CITATION_END -->