**dots.tts** is a 2B-parameter text-to-speech system built on an unusual premise for the current generation of open TTS: there are **no discrete tokens anywhere in the pipeline**. Most modern systems quantize speech into codec tokens, let a language model predict them, then decode back to a waveform. dots.tts keeps the representation continuous end to end, pairing a semantic encoder, an LLM backbone, and an autoregressive flow-matching acoustic head over a **48 kHz AudioVAE**. Skipping quantization removes a lossy step that tends to show up as artifacts in the hardest cases — long sentences, unusual names, emotional delivery. The benchmark results back the design. On **Seed-TTS-Eval**, dots.tts reports word error rates of **0.94% (zh), 1.30% (en), and 6.60% (zh-hard)** with speaker-similarity scores of 81.0 / 77.1 / 79.5, the best average on that suite. On the 24-language **MiniMax multilingual** benchmark it posts the highest average speaker similarity at **83.9**. These are self-reported numbers from the project's technical report, but they are measured on public benchmarks that other open systems also publish against, so they are checkable rather than promotional. What makes the project practically interesting is the **efficiency work layered on top of the base model**. A MeanFlow-distilled variant, `dots.tts-mf`, reaches a real-time factor of **0.15 with 204 ms first-chunk latency** for voice cloning under the optimized inference path, and the team has since shipped `dots.tts-mf-2steps` and `dots.tts-mf-1step` for cloning at just 1-2 function evaluations. **SGLang Omni** now supports all three checkpoint families with continuous batching, streaming PCM, and CUDA-graph decode, hitting 4.64 requests per second on a single H100 at concurrency 16 while holding 1.31% WER. That is the difference between a research artifact and something that can sit behind an API. An August 2026 addition, **dots.tts.edit**, extends the system to instruction-controlled speech editing — changing a word or an inflection in existing audio without regenerating the whole clip, which is the workflow most production voice teams actually need. The honest limitations are the usual ones for a model this size. A 2B autoregressive model with a flow-matching head is not something that runs comfortably on a laptop CPU, and the strongest results assume GPU serving; teams wanting on-device synthesis are better served by smaller distilled systems. The evaluation is also weighted toward Chinese and English, so quality across the 24-language multilingual set is uneven despite the strong average similarity score. Both the code and the checkpoints are **Apache-2.0**, with inference, fine-tuning, and MeanFlow distillation code all in the repository — an unusually complete release rather than weights over the wall.