Fine-tuning Multimodal LLMs for Those Without an A100 Cluster 🦥

The folks at Unsloth have released the beta of Unsloth Studio — a local UI tool for inference and training (Mac, Windows, Linux). Additionally, they have fully implemented support for Gemma 4 from Google. And not just wrapped the API, but manually fixed critical bugs in the original architecture.

To run LoRA for Gemma 4 E2B, you only need 8 GB VRAM. The E4B version fits into 10 GB. If you're aiming for 31B, prepare 22 GB for QLoRA.

And yes, if you have a placeholder instead of a normal graphics card or a MacBook without MLX, Unsloth has prepared a ready-made Colab notebook.
The process is extremely simple:
1️⃣ Run the blocks and Unsloth Studio itself.
2️⃣ Select a model, dataset. Just upload your dataset, choose the layers to train (Gemma 4 now allows separate fine-tuning of text, vision, and audio layers) and go grab a coffee.
3️⃣ Click "Start Training" and watch the progress in real time.
4️⃣ Everything is ready - you can immediately compare the base and fine-tuned versions of the model in the chat.