Breeze TTS 2 is BreezeBlue's open-weight text-to-speech model, released August 25, 2026 with PyTorch inference code on GitHub and safetensors weights on Hugging Face. Built as a single bilingual model covering English and Chinese, it exposes four interaction modes: Voice Clone reproduces a reference speaker's timbre, rhythm, emotion, and style from a short audio sample paired with its exact transcript; Voice Design synthesizes a new voice from a natural-language description with no reference audio; Voice Direction keeps a cloned speaker's identity while a separate instruction string steers tone, emotion, pace, and delivery; and inline vocal events trigger non-speech sounds such as (laugh), (cough), (clears throat), and (sigh) in English text, with bracketed equivalents for Chinese. The inference stack streams audio via a PyTorch runtime with eager execution by default, and an optional --fast-all mode swaps in CUDA Graph-backed execution for the text encoder, backbone, depth decoder, and codec stages, trading a longer cold start for lower latency; per its README, on an NVIDIA H100 with this warmed-up fast path the project reports under 40 ms time-to-first-audio and a 0.32 real-time factor, roughly 3.1x real-time synthesis speed. Eager inference needs about 7.7 GiB of GPU memory (12 GB card recommended), while the full fast path needs about 14.4 GiB (24 GB card recommended); a Docker build script targets H100/Hopper by default, with an A100 flag available. The audio tokenizer is adapted from Alibaba's Qwen3-TTS, also under Apache-2.0. The README states, citing the Artificial Analysis TTS leaderboard, that Breeze TTS 2 ranks first among open-weight models and outperforms some proprietary systems there, though this is the project's own claim, not independently verified. Licensing is split: the inference code is Apache-2.0, but the weights, checkpoints, adapters, derivative models, and self-hosted outputs fall under the separate BreezeBlue Research and Non-Commercial License, and the README states plainly that Apache does not grant commercial rights to the model itself; commercial use of outputs requires a paid subscription to BreezeBlue's hosted platform.