A developer introduced Lure-TTS and Lure-Turbo, two 600M-parameter speech models fine-tuned from Qwen3-TTS using supervised fine-tuning, reinforcement learning with verifiable rewards, and on-policy distillation. Both support inline tags, direction controls, and voice cloning. The developer reports that Turbo runs four times faster than real time and that the full training cost was under $200, but these results have not been independently verified. Training code and an inference engine supporting NVIDIA GPUs, Macs and CPUs were also published, with a Python API, CLI, and an OpenAI-compatible server.