v3.1.0
deepseek-ai/DeepSeek-OCR-2v3.1.0Sep 3, 2026by gabrieldemarmiesse
Upgrading?
AI Summary
This release focuses on performance optimization, specifically improving distillation processes and adding new community models for Hindi and Korean.
Key Highlights
- Added Hindi and Korean community models to the repository
- Optimized `lsd_depth_distill` to process 64 rows in one pass
- Added `--default-voice` option to the serve command
- Replaced beartype with ty for type checking
New Features
- Hindi community model
- Korean 300M community model
- Default voice option for serve command
- Finetune released models into new languages
- Immutable defaults in seanet
Full Release Notes
## What's Changed * Document CPU-only installation (pip pulls ~3 GB of CUDA wheels by default) by @jflaflamme in https://github.com/kyutai-labs/pocket-tts/pull/219 * Speed table: measured H100 rows off local storage by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/250 * lsd_depth_distill: 64 rows in one pass, drop the unused flow multiplier by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/253 * Finetune a released model into a new language by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/254 * Import Annotated from typing by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/256 * Run the training tests in CI, and make zip lengths explicit by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/257 * Run the training tests in CI by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/258 * Fix decimal splitting in sentence chunker by @preciousimo in https://github.com/kyutai-labs/pocket-tts/pull/217 * Split tts_model into text chunking and model-state I/O by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/259 * Immutable defaults in seanet; document the two open() call sites by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/260 * CONTRIBUTING: encourage sharing trained models by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/266 * Add a --default-voice option to the serve command by @gabrieldemarmiesse in https://github.com/kyutai-labs/pocket-tts/pull/271 * README: explain depth+CFG distillation by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/272 * Move text conditioning into modules/text_conditioner.py by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/275 * Use tanh GELU approximation in the backbone FFN by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/278 * Replace beartype with ty for type checking by @gabrieldemarmiesse-ai-agent in https://github.com/kyutai-labs/pocket-tts/pull/283 * Add Hindi community model and configurable terminal punctuation by @Sedherthe in https://github.com/kyutai-labs/pocket-tts/pull/288 * Add "pocket-tts-timestamped" project to README by @dpm63 in https://github.com/kyutai-labs/pocket-tts/pull/290 * Add Korean 300M community model to README by @seastar105 in https://github.com/kyutai-labs/pocket-tts/pull/291 * README: Czech example with a Czech sentence by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/292 ## New Contributors * @jflaflamme made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/219 * @preciousimo made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/217 * @gabrieldemarmiesse-ai-agent made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/263 * @Sedherthe made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/288 * @dpm63 made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/290 * @seastar105 made their first contribution in https://github.com/kyutai-labs/pocket-tts/pull/291 **Full Changelog**: https://github.com/kyutai-labs/pocket-tts/compare/v3.0.2...v3.1.0
Related Repositories
public-apis/public-apis★ 491,198
A collective list of free APIs
NousResearch/hermes-agent★ 256,060
The agent that grows with you
yt-dlp/yt-dlp★ 200,237
A feature-rich command-line audio/video downloader
microsoft/markitdown★ 191,551
Python tool for converting files and office documents to Markdown.
Significant-Gravitas/AutoGPT★ 187,699
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.