IndexTTS-vLLM is an open-source acceleration of the IndexTTS text-to-speech system that re-implements its GPT model inference with vLLM, delivering the same zero-shot voice cloning at much higher speed. The project has about 1,200 stars on GitHub, is licensed under Apache-2.0, and is actively developed in 2026.
IndexTTS, originally from Bilibili’s ecosystem, is known for high-quality zero-shot speech synthesis: give it a short reference clip and it speaks new text in that voice. Its autoregressive GPT stage, however, is the throughput bottleneck. This project rewrites that stage on top of vLLM, the high-performance LLM inference engine. On a single RTX 4090 the real-time factor for IndexTTS 1/1.5 drops from roughly 0.3 to 0.1, and GPT decoding speeds up from about 90 to 280 tokens per second, with comfortable concurrency for serving. Support covers IndexTTS 1, 1.5 and the newer IndexTTS2, including its WebUI and API interfaces.
The repository provides deployment scripts, weight download instructions via ModelScope and Hugging Face, and serving options for both interactive and programmatic use. The synthesis quality itself comes from the underlying IndexTTS models, which are designed for expressive speech; IndexTTS2 additionally targets emotion-aware generation. For teams, the practical appeal is simple: the same voices and quality as IndexTTS, but with serving throughput that makes real-time products feasible.
IndexTTS-vLLM is free and open source under the Apache-2.0 license. You download the model weights and run everything on your own GPU hardware.
IndexTTS-vLLM is for developers deploying IndexTTS in products or services who need production-grade throughput, and for voice AI builders who want fast, self-hosted, clone-capable text-to-speech.