GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
kagakouko
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
kagakouko
Model Tree
Qwen/Qwen3-VL-4B-Instruct
Input Modalities
Output Modalities
Supported Functionality
L1 and L2 establish local state-transition modeling. On L3, verifiable answer rewards supervise the result while same-prefix privileged distillation guides the intermediate reasoning process.
All four checkpoints are listed in the Spatial-Interactor collection.
Use the standard Transformers interface for the base model and load this repository in place of the base identifier:
python
model_id = "kagakouko/Spatial-Interactor-Qwen3-VL-4B"
For video evaluation, preserve chronological frame order and use the frame budget specified by the target benchmark. The paper's main video evaluation uses 32 ordered frames.
The SFT stage uses the reported L1-L2 split of LSI-108K together with the public spatial QA mixture described in the paper. OPD starts from that SFT checkpoint and combines verifiable answer rewards with CoT-only privileged self-distillation on long-horizon video questions. The visual encoder remains frozen while the language model and multimodal projector are updated.
For citation, use the project BibTeX.
This checkpoint is released under Apache-2.0, following the base model license. Users must also comply with licenses and terms governing input datasets and media.