The key difference is that Qwen3-VL-4B-Instruct-Action extends the original tokenizer with an additional 2048 action tokens. Otherwise, it is identical to Qwen3-VL-4B-Instruct—no additional fine-tuning was applied.

To add these tokens, we used the script: starVLA/model/modules/vlm/tools/add_qwen_special_tokens

The model simply expands the token codebook to accommodate action token embeddings, but the underlying VLM weights remain unchanged from the base Instruct model.