Google Gemma 3 now in Unsloth

Gemma 3 and Reasoning is here!

Gemma 3, Google's new multimodal (text + image) model, is now supported in Unsloth! It comes in 1B, 4B, 12B, and 27B sizes, has multilingual support and a 128K context window.

Unsloth also now supports: full finetuning, reasoning (GRPO), 8-bit, pretraining, nearly all models (Mixtral, MOE, Cohere) etc. MultiGPU support is coming very soon.

We’d like to thank the Gemma team for their support and for featuring Unsloth in their release blogpost. Gemma 3 Training free notebooks and guide:

Reasoning (GRPO) Support

Unsloth enables ten times more context lengths and uses 90% less VRAM for GRPO which was the algorithm DeepSeek used to train R1. Unsloth allows you to train your own reasoning model with just 5GB of VRAM using Qwen2.5 (1.5B).

With 24GB VRAM, you’ll be able to transform any model up to 24B in parameters, like Gemma 3 (12B), Llama 3.1 (8B) and Phi-4 (14B) into reasoning models.

---

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!