IndexTTS-2.5
Bilibili's 0.8B zero-shot voice-cloning TTS in five languages, with emotion and speed control, on about 6 GB of VRAM.
In short. IndexTTS-2.5 copies a voice from one short audio clip and uses it to read new text in Chinese, English, Japanese, Spanish or Arabic, even in a language the clip was not in. You can set the emotion separately from the voice, fix how a tricky word is pronounced and speed the speech up or slow it down. It runs on one NVIDIA graphics card with about 6 GB of memory.
Summary
IndexTTS-2.5 is the latest model from the Index team at Bilibili, the Chinese video platform, and its weights were published on 10 August 2026. The technical report is arXiv 2601.03888, which was posted in January 2026. The model is a zero-shot text-to-speech system, so it clones a voice from a single reference clip without any fine-tuning. Compared with IndexTTS-2 it adds Japanese, Spanish and Arabic to Chinese and English, infers faster, adds speaking speed control through a duration factor between 0.5 and 2.0, and improves control over pronunciation written as Chinese Pinyin, English CMU phonemes or Japanese Kana. It keeps the earlier model's cross-lingual voice transfer and its separation of emotion from timbre.
It is an autoregressive model with a GPT backbone of about 0.8 billion parameters, a flow-matching decoder that turns speech tokens into a mel spectrogram, and a BigVGAN vocoder, and it outputs 22.05 kHz audio. Emotion can be set with an eight-number vector for happy, angry, sad, afraid, disgusted, melancholic, surprised and calm, copied from a second reference clip, or described in text through a bundled Qwen-based emotion model that you switch on. The card asks for an NVIDIA GPU, Python 3.10 or 3.11 and roughly 6 GB of VRAM, and the repository is about 5.5 GB. Several helper models, such as w2v-bert-2.0, the MaskGCT semantic codec, CAMPPlus and BigVGAN, download on first run. A vLLM recipe covers production serving, and community GGUF and MLX conversions exist.
Bilibili's own tables on CV3-Eval, with an in-house test set for Arabic, report an average word error rate of 6.75% and a speaker similarity of 73.18 across five languages, against 7.22% and 72.02 for VoxCPM2 and 6.76% and 70.39 for OmniVoice. The README also lists an RL-tuned row with 6.00% and 73.63, and we could not tell which row the published checkpoint matches. On an RTX 4090 it reports a real-time factor of about 0.21 in BF16, against 0.33 for IndexTTS-2 in FP16. All of these are the developer's numbers. The license is Bilibili's own agreement, which has no general ban on commercial use, although the README invites commercial users to contact the team, and it sets thresholds of 100 million monthly active users or RMB 1 billion in revenue and limits using the model to improve other AI models. The card notes that long text is split into segments, so prosody is not modeled across a boundary, and that obtaining the speaker's consent for cloning is the user's job. Hugging Face counted about 29,000 downloads in the 30 days to 11 November 2026.
Good for
- Cloning a voice from a single short clip for narration, dubbing, games and assistants
- Chinese, English, Japanese, Spanish or Arabic speech with cross-lingual voice transfer
- Fine control through an emotion vector or text description, Pinyin, CMU or Kana pronunciation tags, and a speed factor
- A small footprint, since it needs about 6 GB of VRAM and a roughly 5.5 GB download
Watch out for
- The license is custom, so above 100 million monthly users or RMB 1 billion in revenue you need Bilibili's separate approval, and you may not use the model to improve other commercial AI models
- The license text names the model "bilibili indextts2", although the 2.5 card and README point to the same agreement
- It needs an NVIDIA GPU, and the card does not describe CPU or Mac use, although community conversions exist
- Long text is split into pieces, so prosody can reset at the joins, and random emotion sampling lowers voice-cloning fidelity
- The benchmark tables are Bilibili's own, and you are responsible for getting consent from the person whose voice you clone
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of IndexTTS-2.5. 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 |
|---|---|---|---|---|
| IndexTTS-2.5 | 800M | 5.5 GB | 6 GB | Laptop or phone |
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 developer states a minimum of 6 GB of GPU memory.
The license in plain words
Bilibili's license for IndexTTS, which the IndexTTS-2.5 card and the GitHub repository both point to. It lets you use, modify, train on, distribute and build derivatives of the model, worldwide and royalty free. The text has no general ban on commercial use, but a user or affiliate above 100 million monthly active users or RMB 1 billion in annual revenue needs a separate written license.
- Free commercial use only below a revenue cap
- Attribution or model name must be shown
- Acceptable-use policy applies
See also
- Best open-source speech models for transcription and text-to-speech
- Best open-source text-to-speech models in 2026
- Free until you grow: where open licenses draw the line
Frequently asked questions
- What is IndexTTS-2.5?
- IndexTTS-2.5 copies a voice from one short audio clip and uses it to read new text in Chinese, English, Japanese, Spanish or Arabic, even in a language the clip was not in. You can set the emotion separately from the voice, fix how a tricky word is pronounced and speed the speech up or slow it down. It runs on one NVIDIA graphics card with about 6 GB of memory.
- Can I use IndexTTS-2.5 commercially?
- Yes, with conditions. IndexTTS-2.5 is released under the Bilibili Model Use, which allows commercial use but adds extra rules: Free commercial use only below a revenue cap; Attribution or model name must be shown; Acceptable-use policy applies. Read the license text before you ship a product.
- Is IndexTTS-2.5 open source?
- IndexTTS-2.5 is open-weight, not open source in the strict sense. You can download the model files, but the Bilibili Model Use adds conditions that the Open Source Initiative definition does not allow.
- Can I run IndexTTS-2.5 on my own computer?
- Yes. The smallest variant, IndexTTS-2.5, needs about 6 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was IndexTTS-2.5 released?
- IndexTTS-2.5 was released by Bilibili on August 10, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- IndexTTS-2.5 model card (Hugging Face) huggingface.co
- IndexTTS repository, news and evaluation tables (GitHub) github.com
- IndexTTS 2.5 Technical Report (arXiv 2601.03888) arxiv.org
- bilibili Model Use License Agreement (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| VoxCPM2 OpenBMB | 2.29B | – | Apache 2.0 (Permissive) | 8 GB Laptop or phone | |
| OmniVoice Xiaomi | 613M | – | CC BY-NC (Non-commercial) | 3.75 GB Laptop or phone | |
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone | |
| Fun-CosyVoice 3 Tongyi Lab Speech Team | 500M | – | Apache 2.0 (Permissive) | 6.21 GB Laptop or phone |