Today, we’re the #1 trending repo on GitHub and it's all thanks to you! To celebrate, we want to take a moment to welcome anyone whether you’ve been with us from the beginning or just discovered Unsloth! 💖
But first, we wanted to thank each and everyone of you. We’re incredibly grateful for your support whether it’s reading our blogs, engaging with us, or contributing to our repo or just using/sharing Unsloth. Your support truly means the world to us and we wouldn't be here today without you guys!
Also, we're also blown away by how much you guys enjoyed our 1.58-bit Dynamic DeepSeek-R1 and our latest Reasoning release. But this is just the beginning - we’re not slowing down anytime soon!
Here's a little bit about our journey so far:
🦥 Unsloth Highlights
Dec 2023: We launched Unsloth, making training 2x faster and 50% more memory-efficient. We built a custom autograd engine, rewrote all kernels in OpenAI's Triton, including the RoPE embeddings with optimized forward & backward passes and had our own custom Fast Cross Entropy Loss kernel.
Jan 2024: We started uploading models to Hugging Face which has been downloaded by you amazing people millions of times! We also collabed with HF and are so thankful for their continued support. 🤗
Mar 2024: We found & fixed 8 bugs in Google's incredible Gemma models. Thank you Andrej Karpathy for acknowledging our work and the Google team for their support! ❤️
Apr 2024: We introduced Unsloth Gradient Checkpointing which allowed for 4x longer context lengths and 1.7x larger batch sizes. The epic Llama 3 models were released and we enabled the 70B model fit on 48GB of VRAM.
May 2024: We found & fixed many bugs in Llama 3 which greatly improved accuracy for training. The amazing Phi-3 model was released and we Llamafied it.
Oct 2024: We fixed a universal issue in Gradient Accumulation that negatively impacted everyone's training runs (not just for Unsloth but everyone who was training models).
Nov 2024: 2x faster & 70% less VRAM Vision fine-tuning with Llama, Qwen, Pixtral and Llava support. We also found & fixed many Qwen 2.5 bugs.
Dec 2024: We introduced our Unsloth 4-bit Dynamic Quants which greatly increased accuracy compared to standard 4-bit by selectively avoiding quantizing certain layers. And we added support for 13x longer context lengths for Llama 3.3.
Feb 2025: We introduced Reasoning in Unsloth and made GRPO work with QLoRA & LoRA, which previously did not work - allowing home users to train their own reasoning model like R1.
💡 What's Next?
We're always trying to innovate and won't stop just yet. This year, you can expect even more exciting new cool stuff that will make creating your own custom models much easier and more accessible.
And of course - multiGPU & Unsloth Studio are still on the way so don't worry - but we've got to support all models first which is coming pretty soon. Our goal is first and foremost open-source and Unslothing things for you guys, and nothing will ever change that so we hope you follow along this year for even more cool stuff!
💕 Thank you!
A big hug to every single one of you for making Unsloth possible - whether through using, sharing, contributing or engaging with us! 🙏Feel free to join our Reddit page and Discord server for help, to not miss any news or just to show your support! You can also follow us on Twitter and newsletter.
