v3.0.0
deepseek-ai/DeepSeek-OCR-2v3.0.0Aug 25, 2026by gabrieldemarmiesse
Upgrading?
AI Summary
A major release introducing the official training codebase and refactoring internal module structures for better architecture.
Key Highlights
- Training code is now live
- Renamed modules (`mimi_transformer` -> `transformer`)
- Added 24l variant of the English model
- Fixed Docker compose entrypoint args duplication
Breaking Changes
- Module renaming (mimi_transformer -> transformer, transformer -> attention)
- Mimi quantizer ownership changes (decode_from_latent owns quantizer)
New Features
- Training code
- 24l English model variant
Full Release Notes
**The training code is now live at https://github.com/kyutai-labs/pocket-tts/tree/main/training !** EDIT: The 3.0.0 wheel is broken due to a bad build, use 3.0.1. The 3.0.0 wheel has been yanked anyway. ## What's Changed * Fix pocket-tts generate docs by @gabrieldemarmiesse in https://github.com/kyutai-labs/pocket-tts/pull/182 * Add new macOS app links to README by @slaughters85j in https://github.com/kyutai-labs/pocket-tts/pull/193 * Add 'seshat-tts' project to README by @Iheuzio in https://github.com/kyutai-labs/pocket-tts/pull/197 * Update README.md by @joshwhiton in https://github.com/kyutai-labs/pocket-tts/pull/200 * Default the English model's temperature to 0.3 by @manukyutai in https://github.com/kyutai-labs/pocket-tts/pull/223 * Fix audio and state edge cases by @ShiroKSH in https://github.com/kyutai-labs/pocket-tts/pull/210 * Fix docker compose command duplicating entrypoint args by @grigio in https://github.com/kyutai-labs/pocket-tts/pull/233 * docs: correct GPU speedup claim + add CUDA usage notes by @moduvoice in https://github.com/kyutai-labs/pocket-tts/pull/213 * docs: quote pip extras install examples by @nyxst4ck in https://github.com/kyutai-labs/pocket-tts/pull/208 * Rename modules: mimi_transformer -> transformer, transformer -> attention by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/237 * Mimi: decode_from_latent owns quantizer + transposes; encode_to_latent returns [B, T, C] by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/239 * Add the training code by @manukyutai in https://github.com/kyutai-labs/pocket-tts/pull/244 * Add the 24l variant of the english model by @gabrieldemarmiesse in https://github.com/kyutai-labs/pocket-tts/pull/245 ## New Contributors * @slaughters85j made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/193 * @Iheuzio made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/197 * @manukyutai made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/223 * @ShiroKSH made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/210 * @grigio made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/233 * @moduvoice made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/213 * @nyxst4ck made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/208 * @manu-lm made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/237 **Full Changelog**: https://github.com/kyutai-labs/pocket-tts/compare/v2.1.0...v3.0.0
Related Repositories
public-apis/public-apis★ 491,135
A collective list of free APIs
NousResearch/hermes-agent★ 255,963
The agent that grows with you
yt-dlp/yt-dlp★ 200,054
A feature-rich command-line audio/video downloader
microsoft/markitdown★ 191,491
Python tool for converting files and office documents to Markdown.
Significant-Gravitas/AutoGPT★ 187,703
AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the tools, so that you can focus on what matters.