Google's Gemma models get an official JAX toolkit

What it does

This is Google’s own Python package for running and fine-tuning its Gemma family of open-weights LLMs. Built on JAX, it wraps model loading, sampling, and training into a single gm namespace. The same ChatSampler API spans all Gemma versions from 2 through 4, including the newer multi-modal variants that accept images inline.

The interesting bit

Most model releases ship with reference implementations that feel like homework assignments. Here, the unified API across generations is the quietly useful part — you can swap a Gemma 2 text model for a Gemma 4 vision model without rewriting your conversation loop. The library also bakes in LoRA and fine-tuning utilities rather than leaving you to reconstruct them from blog posts.

Key highlights

• First-party JAX implementation from DeepMind, not a community port
• Single ChatSampler API covers text-only and multi-modal (image + text) conversations
• Built-in support for fine-tuning and LoRA, with Colab tutorials
• Runs on CPU, GPU, or TPU; GPU memory requirements start at 8GB for smaller checkpoints
• Model weights downloaded separately via documented checkpoint paths

Caveats

• Gemma 4 technical report is listed as “coming soon,” so capabilities claims for the newest model are currently unverified
• GPU memory jumps to 24GB+ for 7B-parameter checkpoints, which may exclude consumer hardware for larger variants

Verdict

Worth a look if you’re already in the JAX ecosystem and want Google’s open models without Hugging Face indirection. Skip it if you’re committed to PyTorch or need production serving infrastructure — this is a research and experimentation toolkit, not a deployment platform.