Llasa 3B

Llasa 3B is the 3-billion-parameter checkpoint of LLaSA, the February 2025 HKUST project that reframed text-to-speech as plain next-token prediction. Initialized from LLaMA with an expanded vocabulary of XCodec2 speech tokens (65,536 codebook entries), it trains TTS exactly like training an LLM — no separate diffusion stage, no multi-model cascade — on 250,000 hours of Chinese-English speech. The paper (arXiv 2502.04128) systematically studies scaling both train-time and inference-time compute for speech, showing inference-time verifier search improves emotional expressiveness, timbre consistency, and content accuracy.

Checkpoints and training code for the 1B, 3B, and 8B variants were released publicly, and the 3B hits the practical middle ground for self-hosting. The honest picture: the authors note the minimalist design does not always match heavily customized TTS systems on raw naturalness, inference-time scaling adds latency and complexity, and community fine-tunes (like the Japanese Galgame variants) carry their own non-commercial licenses — check the specific checkpoint's terms before shipping anything.