Pocket TTS English 1.9M REFL 2026-10
Pocket TTS English 1.9M REFL 2026-10 is a downloadable text-to-speech checkpoint in the Pocket TTS family. Its official configuration identifies the trained weights and the components needed to generate english speech.
Research release
The configuration identifies this release
The configuration defines an 8-layer backbone and a reduced sampling head and decoder. It provides separate weight paths for the standard and no-voice-cloning packages. The reward-fine-tuned checkpoint has its own release path.Pocket TTS english_2m_refl_2026-10: official release configuration
REFL adds reward fine-tuning
The REFL version adds reward fine-tuning after distillation.Pocket TTS: smaller models and training method
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | pocket-tts:english_2m_refl_2026-10kyutai/pocket-tts/languages/english_2m_refl_2026-10kyutai/pocket-tts-without-voice-cloning/languages/english_2m_refl_2026-10 |
| Backbone depth | 8 transformer layers.Pocket TTS english_2m_refl_2026-10: official release configuration |
| Released parameter count | 1.9M including decoder.Pocket TTS: smaller models and training method |
| Weight license | CC BY 4.0.pocket-tts-without-voice-cloning: Kyutai model card |
| Voice-cloning weights | Repository acceptance required.pocket-tts: Kyutai model card |
Related models
- Pocket TTS English 2026-10
Kyutai ยท Pocket TTS
An October 2026 English speech checkpoint that retrains September revision 2 with a corrected training loader.
Launch date not confirmedDownloadable