Check out Unsloth Studio it provides MTP support now which 2x the token generation speed with no loss of accuracy: https://unsloth.ai/docs/models/qwen3.6#mtp-guide