Text-only checkpoint extracted from the language-model tower of Qwen/Qwen3-VL-2B-Thinking and rewritten as a standard model_type: qwen3 / Qwen3ForCausalLM checkpoint: the vision and aligner modules are dropped and model.language_model.* keys are renamed to model.*.

Intended as a text-SFT base so trainers instantiate a plain Qwen3 causal LM instead of the full Qwen3VLForConditionalGeneration class.

Built with the deepswe converters sft/qwen3/scripts/prepare_qwen3_vl_text_checkpoint.py + save_qwen3_vl_text_as_qwen3_checkpoint.py.