whatmodel.sh / tts / LongCat-AudioDiT-3.5B

LongCat-AudioDiT-3.5B

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

LongCat-AudioDiT-3.5B is an open-weights text-to-speech model from Meituan, released March 2026. Its published size is 3.5B parameters. Licence: mit.

Catalogue identity notes

Meituan LongCat team's diffusion TTS in the waveform latent space, exact checkpoint meituan-longcat/LongCat-AudioDiT-3.5B @ 5d3e6dc007872d756c3a69b44951598d7d9bdf6d (MIT; created 2026-03-30). The 3.5B row of the LongCat-AudioDiT Seed-benchmark table (card and arXiv:2603.29339v1 Table 1). Distinct from the LongCat LLMs and LongCat-Audio-Codec.

Model information sources

Release date ↗ · Weights licence ↗ · Parameter count ↗

Maker's size label 'LongCat-AudioDiT-3.5B' (diffusion backbone). HF safetensors header total 3,833,985,217 F32 parameters for the single model.safetensors, which also holds the Wav-VAE and UMT5-base text encoder.

65.8±2.0#10 of 118 text-to-speech modelsfrom 10 benchmark results across 2 sources

Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.

$ how it compares

The text-to-speech models ranked either side of LongCat-AudioDiT-3.5B, out of 148 with a score.

#modelscorewhereRAMreleased
14dots.tts Base (Pretrain) 2B dots.tts67±2comparison only2026-06
15MiniMax Speech 2.8 Turbo MiniMax66±3api $60.00/M chars2026-02
16LongCat-AudioDiT-3.5B Meituan66±2local~15GB2026-03
17LongCat-AudioDiT-1B Meituan66±4local~5.7GB2026-03
18OmniVoice k2-fsa (Xiaomi)66±2local~3.3GB2026-03

see all text-to-speech models · find the ones that fit your machine

$ what it's good at

voice cloning▁·▁▁▁▁▁▁▇█▃▃ #4 of 37 · 43% of its score · 6 tests

Beats Gemini 3.1 Flash TTS Preview, dots.tts MF 2B (MeanFlow, NFE 4), FireRedTTS3 (base) here, though they rank higher overall. Trails dots.tts SOAR 2B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Seed-TTS test-zh · speaker similarity (ERes2Net SV)80.6% ✎█·····██···█ beats 1/3medium · avg 58%▇▇▇▇ 11%typical
Seed-TTS test-en · speaker similarity (ERes2Net SV)77.1% ✎█····█··█··█ beats 1/3medium · avg 54%▇▇▇▇ 11%typical
Seed-TTS test-en · speaker similarity (WavLM-large SV) lab78.6% ✎▁·▂▂▂█▆▃▅··▂ beats 27/30medium · avg 53%▇▇▇ 8.3%at its level · small run
Seed-TTS test-zh · speaker similarity (WavLM-large SV) lab81.8% ✎▁·▂··▂·█▄·▃▄ beats 25/25medium · avg 62%▇▇▇ 7.3%small run
Seed-TTS test-en · Whisper large-v3 WER lab1.5% ✎▁···▂▂▂▂█▆▄▂ beats 28/35avg 3.6%▇ 3.3%blunt · small run
Seed-TTS test-zh · Paraformer CER lab1.1% ✎▁·▂▁▁▁▂▃▅█·▂ beats 16/30avg 2.1%▇ 2.5%blunt · small run
other tests4 not in a task · 55% of its score
testresultvs the others in its rundifficultycountswhy
CV3-Eval hard-en · WER8.5% ✎█····█·█████ beats 1/6avg 26%▇▇▇▇▇▇▇▇ 19%sharp
CV3-Eval hard-en · DNSMOS (predicted quality)3.84 / 5 ✎█·▄·▄▄·▄···▄ beats 3/6medium · avg 2.98 / 5▇▇▇▇▇▇ 15%at its level
CV3-Eval hard-zh · CER9.2% ✎▄▄▄···▄··▄·█ beats 3/6avg 23%▇▇▇▇▇ 14%typical
CV3-Eval hard-en · speaker similarity (ERes2Net SV)73.9% ✎▄·····▄▄·▄▄█ beats 4/6medium · avg 49%▇▇▇ 7.0%at its level

$ where to run it

Other ways to run it

  1. Run it as published · needs ~15.34 GB · from huggingface.co/meituan-longcat
    git clone https://github.com/meituan-longcat/LongCat-AudioDiT.git && cd LongCat-AudioDiT && pip install -r requirements.txt && python inference.py --text "Today is a sunny day, let's go for a walk in the park." --output_audio output.wav --model_dir meituan-longcat/LongCat-AudioDiT-3.5B

Weights and documentation: huggingface.co/meituan-longcat ↗

$ how the score was worked out · Qwen-Audio-3.0-TTS technical… 77%, meituan-longcat/LongCat-Audio… 21%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Qwen-Audio-3.0-TTS technical report6▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 77%CC BY 4.0
meituan-longcat/LongCat-AudioDiT-3.5B model card (Seed benchmark table) · lab-reported4▇▇▇▇▇▇ 21%MIT
prior · pull toward the middle▇ 2%
weighted average= 65.8

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json