LongCat-AudioDiT-3.5B

LongCat-AudioDiT-3.5B

Turns Chinese or English text into speech with a diffusion model that works directly on the waveform latent space, and can clone a voice from a reference sample you have the right to use.

Made by meituan-longcat on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: LongCat-AudioDiT-3.5B is a text-to-speech model from Meituan's LongCat team. It generates audio with a diffusion backbone that works on a waveform latent space, so it skips intermediate steps such as mel-spectrograms.

The pipeline needs only two parts, a waveform variational autoencoder and the diffusion backbone. The card lists Chinese and English as its languages and reports zero-shot voice cloning results.

Who it suits: Speech researchers and developers who want an open text-to-speech model with published code and weights. Voice cloning is a feature here, so use only voices you have the right to use.

Industry | Who uses it | How they use it | Example
Research | Speech researchers | Study diffusion text-to-speech in a waveform latent space | Compare its scores with Seed-TTS on the Seed benchmark
Software | App developers | Add synthetic voices in Chinese or English | Read short app messages aloud
Media | Content producers | Narrate scripts without recording every line | Turn a written script into spoken audio

People who want a ready-made voice app can skip it, since the page offers model files and research code, not an app.

Where to use it:

• Python scripts, using the Hugging Face compatible implementation in the repository
• Windows, macOS or Linux with Python, for downloading the files
• Chinese and English speech work

Use case | Who | What they get back
Generate speech from text | Developers | Spoken audio in Chinese or English
Clone a voice from a reference | Researchers and producers | Speech in a similar voice, if you hold the rights to it
Benchmark text-to-speech | Researchers | Speaker similarity figures to compare with other systems

How to set it up: Plan for a large download first, since the weights are about 15.3 GB in F32.

• Python and pip, for the Hugging Face download tools
• Disk space for the model files, about 15.3 GB in F32 by Hugging Face's count
• The repository's inference scripts, which the card says come with the model

1. Install the huggingface_hub package.
2. Download the model files with the hf command.
3. Follow the card's inference scripts to generate speech.

pip install -U huggingface_hub
hf download meituan-longcat/LongCat-AudioDiT-3.5B

How to use it: The card says the repository provides the model definition, weight conversion and inference scripts. Run those scripts with your text. For voice cloning you also give a reference recording of the speaker. The usage part of the card is cut off here, so check the model page for the exact script names.

Example: You give it a short Chinese or English sentence and, for cloning, a reference recording you have permission to use. It returns an audio file of the sentence spoken in a similar voice.

For newcomers and experts: Think of it as a model that writes a voice recording from text. A good first try is to download the files, open the card and run one of its inference scripts on a single sentence.

Underneath, a waveform VAE compresses audio into a latent space and a diffusion backbone generates in it, without autoregressive decoding. The authors fixed a training-inference mismatch and replaced classifier-free guidance with adaptive projection guidance. They also report that a better VAE reconstruction does not always give better speech.

Text and a reference voice go in
The diffusion backbone works in latent space
The waveform VAE decodes the latent
Speech audio comes out

Good to know: The figures below are the authors' own results on the Seed benchmark, comparing speaker similarity with Seed-TTS.

Title: Speaker similarity, higher is better
Seed-TTS on Seed-ZH: 0.809 SIM
LongCat-TTS-3.5B on Seed-ZH: 0.818 SIM
Seed-TTS on Seed-Hard: 0.776 SIM
LongCat-TTS-3.5B on Seed-Hard: 0.797 SIM

• Size: the name says 3.5B, and Hugging Face's record counts 3.8B parameters in Safetensors
• Variants: the card also links a smaller 1B version, and Hugging Face shows 2 quantized copies by other people
• License: MIT
• Cost to run: no API key or paid service is named
• Data it reaches: the text and reference audio you give it, plus the model files you download from Hugging Face
• Not for: the card names no restrictions beyond its license, but cloning a voice needs the speaker's permission

On Hugging Face, 13 November 2026

• Parameters: 3.8B
• Weights: about 15.3 GB in F32; about 1.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,026
• Likes: 83
• License: mit
• Languages: zh, en
• Quantized versions: 2
• Tags: safetensors, audiodit

Links

• Model: https://huggingface.co/meituan-longcat/LongCat-AudioDiT-3.5B
• Files: https://huggingface.co/meituan-longcat/LongCat-AudioDiT-3.5B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download meituan-longcat/LongCat-AudioDiT-3.5B

Then follow the usage in its model card. Nothing in this kit runs by itself.