Unsloth 支持 Google Gemma 3n 微调并新增 TTS 与 DeepSeek-R1-0528 支持
Unsloth - Google Gemma 3n & TTS
Unsloth 宣布支持 Google 新模型 Gemma 3n 的运行与微调，该模型支持文本、图像、视频和音频，上下文为 32K，官方已上传 Dynamic quants、GGUF 并发布微调 notebook。
Unsloth 一次性放出 Gemma 3n 微调、TTS 支持和 DeepSeek-R1-0528 量化方案，可据此判断本地微调与部署的可行路径。
Happy July! 🦥 We’re excited to introduce lots of new updates for Unsloth:
✨ Gemma 3n
Run and fine-tune Google’s new Gemma 3n models. Gemma 3n supports text, image, video, and audio with 32K context. We’ve uploaded Dynamic quants, GGUFs, and released fine-tuning notebooks.
🦥 Unsloth Updates
We’ve released a new GitHub repo featuring 100+ Unsloth notebooks
Qwen3 Advanced GRPO notebook with our new proximity scoring reward function and priming stage to skip GRPO format learning
Unsloth now supports the latest vLLM, TRL, and Transformers.
New Reinforcement Learning (RL) and LoRA hyperparameter guide
🔊 Text-to-Speech (TTS)
Unsloth supports TTS and Speech-to-text (STT) models including Sesame-CSM, Orpheus, OpenAI’s Whisper. We also made extensive notebooks for TTS and a Guide.
🐳 DeepSeek-R1-0528
DeepSeek’s R1 update is here. Fine-tune DeepSeek-R1-0528-Qwen3 with GRPO. Our notebook features a new reward function which increases multilingual response rates by 40%.
Our Dynamic 1-bit GGUFs shrink the full 715GB model to 162GB (-80% size).
来源：Unsloth Blog · unslothai.substack.com