X2Streaming-TTS is the reference implementation of "X2Streaming-TTS: Causal Token-Level Text-to-Speech from Streaming Text with Speech-State Inheritance" (arXiv:2608.18661): causal commitment holds an ambiguous span like the "3" in "He finished 3" until a later token resolves its reading, then closes segments through capacity-adaptive, punctuation-aware segmentation before the acoustic budget runs out; causal speech-state inheritance warm-starts the next segment's Code2Wav state and carries the trailing H=4 Talker states through bounded causal-prior attention across the boundary. The method runs on X Square Robot's separate Qwen3TTS-Streaming engine, referenced only as a pinned git submodule; this repo holds the method code, its X2StreamingPolicy facade, and two lifecycle patches connecting to it — not the engine's TensorRT export, scheduler or gateways. Two Apache-2.0 checkpoints are published on Hugging Face: X2Streaming-TTS-1.7B, a ~4.52 GB CustomVoice model fine-tuned from Qwen3-TTS-12Hz-1.7B-Base under speaker robot_service_v1, and X2-NativeCursor-Qwen3TTS-12Hz, a 2.0M-parameter, 8.2 MB observer head from the companion paper "X2-NativeCursor" (arXiv:2609.09677) that reads codebook-0 tokens to publish a monotonic reading-progress cursor before waveform decoding; enabling it depends on the engine's own release checks. Paper benchmarks report a boundary pitch discontinuity of 22.61 Hz against 31.68 Hz for FireRedTTS-2, the best chunk-level comparator, and a median time to first audio token of 15.8 ms for one request, 260.8 ms at 128 concurrent requests, on one RTX 5090. A newer quickstart.sh launcher, separate from that submodule-plus-patch workflow, downloads both checkpoints, builds the TensorRT engines in Docker on a Linux NVIDIA GPU machine, and can expose a temporary Cloudflare WSS endpoint (--public) for the browser Demo; its docs note that a complete cold GPU build of the launcher has not yet been validated. The code is MIT and the Qwen3-TTS base is Apache-2.0. The project is explicitly pre-release: the extraction and patch series were exercised on a real custom-1.7b TensorRT checkpoint on an RTX 4090 D, with broader validation still in progress.