Unsloth - Qwen3 & Dynamic 2.0
Introducing Dynamic 2.0 quants and Qwen3.
We’re excited to introduce Unsloth Dynamic 2.0 which outperforms leading quantization methods and sets new benchmarks for 5-shot MMLU and KL Divergence. This means you can now run + fine-tune quantized LLMs with minimal accuracy loss!
The new Qwen3 models deliver advancements in reasoning, instruction-following and multilingual support.
You can fine-tune Qwen3 up to 8x longer context lengths with Unsloth compared to all setups with FA2 on a 24GB GPU. Qwen3-30B-A3B comfortably fits on 17.5GB VRAM. We released a Colab notebook for Qwen3 (14B) here.
We uploaded Qwen3 GGUFs using our Unsloth Dynamic 2.0 method and also uploaded Qwen3 with native 128K context windows (up from 40K).
Other Highlights:
Meta released Llama 4 in Scout and Maverick variants. You can run them locally with our Dynamic 2.0 GGUFs.
We collaborated with Meta on a Synthetic Datasets notebook.
Microsoft released their new reasoning models for Phi-4 which you can now run or fine-tune with Unsloth.
Thank you! If this email landed in your "Promotions" tab, drag it to your "Primary" inbox and ⭐️Star it to ensure you don’t miss an update from us. We appreciate your support!