- Fixed a critical vocabulary misalignment where text inputs triggered continuous sequence repetition (e.g., infinite numeric/character streams).
- Switched backend text processing to the official Qwen3.5-4B base tokenizer to eliminate multimodal vision artifacts and enforce pure text inference.
- Stabilized reasoning parameters; the model now correctly triggers its `</thinker>` block seamlessly.
- Re-quantized to Q6_K with native llama.cpp protocols for optimized consumer-grade GPU execution.
Delete generation_config.json with huggingface_hub
