# BreezeBlue releases Breeze-TTS-2, a 3B English and Chinese TTS, non-commercial

> Breeze-TTS-2 is a roughly 3B bilingual TTS with cloning, voice design and streaming; weights under a non-commercial licence.

Published 2026-08-25

Canonical: https://shipped.anth.us/open-weights/breezeblue-breeze-tts-2/

BreezeBlue released the weights and PyTorch inference code for Breeze-TTS-2 on 25 August 2026, per its card. It is an open-weight model of about 3B parameters covering English and Chinese in one model. The card lists Voice Clone (reference audio plus exact transcript), Voice Design from a text description, Voice Direction for steering tone and pace while cloning, inline vocal events such as "(laugh)", and a streaming API returning 24 kHz mono PCM.

The publisher claims a #1 rank among open-weight models on the Artificial Analysis TTS leaderboard; that is a vendor claim and the leaderboard was not opened for this item. It also claims under 40 ms to first audio and a real-time factor of 0.32 on an H100 with the fast path, and about 7.7 GiB of GPU memory in eager mode. A 12 GB GPU is the minimum recommended, on Linux with NVIDIA CUDA.

Licensing: source code is Apache 2.0, while weights, derivatives and self-hosted outputs fall under the BreezeBlue Research and Non-Commercial License. Hosted-platform outputs can be used commercially with a paid plan. The tokenizer is based on Qwen3-TTS.

Open weights with a subscription next to them.

## Sources
- [BreezeBlue model card on Hugging Face](https://huggingface.co/BreezeBlue/Breeze-TTS-2)
- [Hugging Face repo metadata (creation date)](https://huggingface.co/api/models/BreezeBlue/Breeze-TTS-2)

## Editor's first pass (AI-assisted)
- Real 4: Weights and PyTorch inference code are released; self-hosted use is non-commercial.
- Useful 3.5: Cloning, design and a streaming API in one model; two languages only.
- Notable 3.5: The card claims the top open-weight rank on a third-party leaderboard, which I did not open.
- Novel 3: Feature set matches its peers; it reuses a Qwen3-TTS-based tokenizer.