MOSS-TTS

OpenMOSS's 8B and 4B Apache 2.0 text-to-speech models with 31 languages, voice cloning, Pinyin and IPA input and exact pauses.

In short. MOSS-TTS reads text aloud in 31 languages and can copy a voice from a recording. You can spell out pronunciation with Pinyin or IPA, set how long a sentence should take, and add pauses of an exact length. The weights are free to use in a business.

Summary

MOSS-TTS is the flagship of the MOSS-TTS family from MOSI.AI and the OpenMOSS team. The v1.5 model has 8B parameters and, according to the README's news log, came out on 26 May 2026, continued from MOSS-TTS 1.0. A second checkpoint, Local Transformer v1.5, followed on 18 June 2026. It has 4.55B parameters, swaps the Qwen3-1.7B backbone of the earlier Local model for Qwen3-4B, and uses the MOSS-Audio-Tokenizer-v2 codec to write native 48 kHz stereo audio. Both cover 31 languages, the 20 of version 1.0 plus Cantonese, Dutch, Finnish, Hindi, Macedonian, Malay, Romanian, Swahili, Tagalog, Thai and Vietnamese. They clone a voice zero-shot, generate long passages, accept Pinyin or IPA for exact pronunciation, set duration by token count and take inline pauses such as [pause 3.2s].

The cards say to set a language tag whenever you know the language, because without one v1.5 can improve some languages and regress slightly on others compared with 1.0. They give no new benchmark table and point back to the 1.0 results. In OpenMOSS's own Seed-TTS-eval table for the 1.0 models, the 8B model has an English word error rate of 1.84% and a Chinese character error rate of 1.37%, while Qwen3-TTS 1.7B lists 1.5% and 1.33%. The 1.7B Local model has the highest speaker similarity among the open models in that table, at 73.28% for English and 79.62% for Chinese. These are the developer's numbers for the earlier models, so they are a guide and not a ranking of v1.5.

The weights are Apache 2.0, at 17.0 GB for the 8B model and 9.1 GB for the 4B in BF16, and the cards recommend Transformers 5 with custom code loaded through trust_remote_code. OpenMOSS also documents a PyTorch-free path with llama.cpp and an ONNX codec. For the 1.0 model it publishes GGUF weights and says the 8B fits on an 8 GB GPU that way. vLLM-Omni supports the family, and SGLang-Omni added day-0 support for Local Transformer v1.5 with an OpenAI-compatible speech endpoint. In the 30 days to 11 November 2026, Hugging Face counted about 78,000 downloads for MOSS-TTS-v1.5 and 82,000 for Local Transformer v1.5. The same team's MOSS-TTS-Nano and MOSS-TTSD are listed as separate models here.

Good for

  • Narration, audiobooks and dubbing in many languages, with 31 on the list and long passages in one request.
  • Chinese and English texts where pronunciation must be exact, using Pinyin or IPA input.
  • Timed audio, with token-level duration control and inline pauses such as [pause 3.2s].
  • Serving behind an OpenAI-style speech endpoint with SGLang-Omni or vLLM-Omni.

Watch out for

  • The v1.5 cards publish no new benchmark, and the 1.0 numbers they point to are the developer's own.
  • Without a language tag, v1.5 can be slightly worse than 1.0 in some languages, so set the tag whenever you know the language.
  • The setup needs Transformers 5 and trust_remote_code, and the 8B model is 17.0 GB in BF16, so plan for a GPU with a good margin above that.
  • Clone a real person's voice only with that person's consent.

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of MOSS-TTS. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Local Transformer v1.5 (4B) BF16 4.55B 9.1 GB 10 GB Laptop or phone
v1.5 Delay (8B) BF16 8.49B 17 GB 20 GB Gaming GPU

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is MOSS-TTS?
MOSS-TTS reads text aloud in 31 languages and can copy a voice from a recording. You can spell out pronunciation with Pinyin or IPA, set how long a sentence should take, and add pauses of an exact length. The weights are free to use in a business.
Can I use MOSS-TTS commercially?
Yes. MOSS-TTS is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is MOSS-TTS open source?
MOSS-TTS is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run MOSS-TTS on my own computer?
Yes. The smallest variant, Local Transformer v1.5 (4B), needs about 10 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was MOSS-TTS released?
MOSS-TTS was released by OpenMOSS on May 26, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
MOSS-TTS-Nano OpenMOSS 100M – Apache 2.0 (Permissive) 0.276 GB Laptop or phone
MOSS-TTSD OpenMOSS 8.36B – Apache 2.0 (Permissive) 19 GB Gaming GPU
Qwen3-TTS Alibaba Qwen 910M – 1.92B – Apache 2.0 (Permissive) 3.68 GB Laptop or phone
VoxCPM2 OpenBMB 2.29B – Apache 2.0 (Permissive) 8 GB Laptop or phone