Excited to share Penzai, a JAX research toolkit from Google DeepMind for building, editing, and visualizing neural networks! Penzai makes it easy to see model internals and lets you inject custom logic anywhere. Check it out on GitHub:

Penzai integrates seamlessly with @GoogleColab and the JAX ecosystem. It represents models as legible, editable data structures, to help researchers understand and modify them after they are trained. Built with support from @DougalMaclaurin, @dtarlow2, and @hugo_larochelle!

Want to get started? Penzai's documentation ( includes guided tutorials that show how to visualize, analyze, and fine-tune the Gemma models in Colab. Interpreting attention heads: Low-rank finetuning:

Penzai's goal is to reduce the barrier of entry for research on understanding pretrained neural networks and steering their behaviors, and to make it easier for researchers to quickly try out new ideas. I'm excited to see what the community can do with it!

@GoogleDeepMind god i wish i used jax

@GoogleDeepMind Is there something similar for @PyTorch ?

@froystig @GoogleDeepMind Wow this is beautiful!

@GoogleDeepMind Dang, I need more hours in the day to check this out. Amazing work.

@GoogleDeepMind Looks really cool! Qq, can you port existing flax models weights to this, without the initial model being written in penzai.nn?

@GoogleDeepMind Flax models are partially supported, but they are harder to visualize due to how Flax represents submodules. But you can write a Penzai version that uses the same parameters. See the "Gemma from Scratch" tutorial for more information on how to port models from Flax to Penzai!

@GoogleDeepMind So cool!