MTP gets lost in quantizing of Qwen3 Next. The implementation is currently overly dependent on the model being DeepSeek v3/r1