marimo

Full Stack AI Developer Meetup at NVIDIA GTC Attending NVIDIA GTC? Join the AI Innovators happy hour to talk shop on advanced AI workflows from reinforcement learning and agent evaluations to next-generation notebook environments. We will dive into how teams use the latest AI workflows to build more reliable agents, everything from personal assistants to robots, all on a purpose-built AI cloud.

If you're going to NVIDIA GTC, come hang with us! marimo is joining CoreWeave, @weightsbiases.bsky.social, Anyscale, Monolith, and @openpipe.bsky.social. We'd love to hear what you're building.

📅 March 17th, 7:30pm PT | San Jose

We'll see you there: → wandb.ai/site/resour...

Prompt Repetition Improves Non-Reasoning LLMs View 2 comments: The prompt repetition technique shows good gains for options first formatting but only modest gains for the questions first formatting. I would like more details on how the options-first case was impl...

Read the paper on @askalphaxiv: www.alphaxiv.org/abs/2512.14982

LLMs Work Better When You Repeat—But Why? When you convert a marimo notebook you can expect a 10x lighter file (and we're being generous). But that's not even the best news! The best news is that we'...

Researchers at Google just released a new trick for increasing the performance for LLMs, and it's much simpler than you might think: just repeat the prompt!

In our latest YouTube video, we try out this technique in a marimo notebook.

www.youtube.com/watch?v=j2v...

mo.status.progress_bar is now thread-safe, making it possible for multiple mo.Thread objects (started in the same cell) to update a single progress bar. This lets you communicate progress to the user when multiple threads are sharding work.

Try it on molab: molab.marimo.io/github/mari...

PyTorch nn.Module instances now render as collapsible HTML trees with color-coded layer categories.

Try it on molab: molab.marimo.io/notebooks/n...

matrix - marimo-team/gallery-examples View and run matrix.py from marimo-team/gallery-examples on GitHub as an interactive Python notebook with marimo.

mo.ui.matrix provides interactive numeric grid inputs. Initialize with a nested list or NumPy array and get a reactive element that updates as the user edits cells.


Try it on molab: molab.marimo.io/github/mari...

visualizing-embeddings - marimo-team/gallery-examples View and run visualizing-embeddings.py from marimo-team/gallery-examples on GitHub as an interactive Python notebook with marimo.

mo.ui.matplotlib adds reactive box and lasso selection to Matplotlib scatter plots. Make a selection on a scatter plot and your selection is automatically sent back to Python.

Try it on molab: molab.marimo.io/github/mari...

marimo 0.20.0 Reactive Matplotlib selections, matrix inputs, rich PyTorch display, and more

Release blog: marimo.io/blog/marimo...

Our latest release brings richer interactivity to machine learning and scientific computing workflows, including reactive @matplotlib plots, a matrix UI element, a native rich visualizer for @PyTorch modules, and a thread-safe progress bar for parallelizing work.

(1/n)

Turns out migrating from Jupyter to marimo is so easy that's it's fun:

youtu.be/Bi-ZfPRg08w

Run marimo notebooks on any cloud with SkyPilot marimo notebooks are reproducible Python programs. SkyPilot runs Python anywhere. Together, they take notebooks from laptop to cloud with zero rewrites.

marimo notebooks are reproducible Python programs. @skypilot_org runs Python anywhere. Together, they take notebooks from laptop to cloud with zero rewrites.

Learn more at our latest blog: marimo.io/blog/skypilot

I want to understand a paper, do LLMs actually save me time? Do you know that feeling of seeing an interesting paper and wondering if it really works? Maybe try it out on a new task? Back in the old days you could say ...

Our latest video is the impact AI can have on education and learning, but the angle might not be what you expect:

www.youtube.com/watch?v=J8u...

Reinventing notebooks as reusable Python programs Designing a Python notebook that blends the best parts of interactive computing with the sanity of code

marimo is the only open-source Python notebook that doubles as a Python script and module. This yields a great developer experience for humans and agents alike.

marimo.io/blog/python...

With skills like these, companies ranging from the F500 to small startups are replacing both Jupyter notebooks and legacy SaaS tools like Retool, Hex, and Tableau with marimo en masse.

Skills - marimo The next generation of Python notebooks

docs.marimo.io/guides/gene...

Extending marimo with General Agent Skills Thanks to skills.sh, we now have a general way to define skills and to share them across all major coding agents. This video shows some skills that we've imp...

Announcing our skills repo, which contains curated instructions for agents to write marimo notebooks: convert Jupyter to marimo, write high quality notebooks, write custom UI elements from scratch, and more.

www.youtube.com/watch?v=-MZ...

Who's Harry Potter? Approximate Unlearning in LLMs View recent discussion. Abstract: Large language models (LLMs) are trained on massive internet corpora that often contain copyrighted content. This poses legal and ethical challenges for the developers and users of these models, as well as the original authors and publishers. In this paper, we propose a novel technique for unlearning a subset of the training data from a LLM, without having to retrain it from scratch. We evaluate our technique on the task of unlearning the Harry Potter books from the Llama2-7b model (a generative language model recently open-sourced by Meta). While the model took over 184K GPU-hours to pretrain, we show that in about 1 GPU hour of finetuning, we effectively erase the model's ability to generate or recall Harry Potter-related content, while its performance on common benchmarks (such as Winogrande, Hellaswag, arc, boolq and piqa) remains almost unaffected. We make our fine-tuned model publicly available on HuggingFace for community evaluation. To the be

Open the notebook in molab: molab.marimo.io/notebooks/n...

Read the paper on @askalphaxiv: www.alphaxiv.org/abs/2310.02238

Can an LLM unlearn Harry Potter? Is it possible to make an LLM forget about a topic? It turns out: yes! And it involves overfitting on the topic that you are interested in! Full explainer ca...

Is it possible to make an LLM forget about a topic?

Yes, it turns out!

Learn how in our latest video:
www.youtube.com/watch?v=HwA...

matplotlib just got another interactive upgrade! When we made the first interactive PuckChart we didn't realize that a box/lasso utility on a scatter chart would be a better idea. But thanks to a comment on...

Select points in a @matplotlib scatter plot, and get them back immediately in Python — only in marimo

youtu.be/c-nasIVbAoM

Release 0.19.8 · marimo-team/marimo This release introduces notebook galleries with OpenGraph previews, along with anywidget improvements, AI enhancements, and many bug fixes. ⭐️ Highlights Serve a gallery of notebooks You can now se...

Release notes: github.com/marimo-team...

PDF export. Export notebooks to PDFs from the CLI or marimo editor, with HTML outputs captured as images.

Multi-modal AI – you can now attach images when using "Generate with AI". Go from whiteboard scrawlings to Tikz diagrams instantly

Customize gallery cards with OpenGraph metadata, letting you express yourself with rich thumbnails and descriptions

Just shipped: marimo v0.19.8! Like all our releases, this one is packed. Highlights:

Serve a gallery of notebooks with `marimo run`. Gallery cards can be customized with OpenGraph metadata, letting you showcase your work with custom thumbnails. (1/n)

GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization View recent discussion. Abstract: As language models become increasingly capable, users expect them to provide not only accurate responses but also behaviors aligned with diverse human preferences across a variety of scenarios. To achieve this, Reinforcement learning (RL) pipelines have begun incorporating multiple rewards, each capturing a distinct preference, to guide models toward these desired behaviors. However, recent work has defaulted to apply Group Relative Policy Optimization (GRPO) under multi-reward setting without examining its suitability. In this paper, we demonstrate that directly applying GRPO to normalize distinct rollout reward combinations causes them to collapse into identical advantage values, reducing the resolution of the training signal and resulting in suboptimal convergence and, in some cases, early training failure. We then introduce Group reward-Decoupled Normalization Policy Optimization (GDPO), a new policy optimization method to resolve these issues by d

For the best reading experience, check out the paper on @askalphaxiv!

www.alphaxiv.org/abs/2601.05242

GDPO vs GRPO - molab marimo notebook by Vincent in marimo team

Notebook: molab.marimo.io/notebooks/n...

From alphaXiv to interactive Python notebook There's a new hobby for ML people: turning papers into elaborate notebooks. Widgets and all! This video explores one example of this indepth. 00:00 Alphaxiv ...

GDPO is a very simple idea. But you might not know that from reading the paper!

Watch our latest video to see how to go from research papers to interactive explainer notebooks — using LLMs to make widgets that bring equations to life.

www.youtube.com/watch?v=Ybz...

marimo's pure-Python file format makes it easy for agents to author notebooks. Watch this video to learn how pairing marimo with a fast model — like @cursor_ai's Conductor — can rapidly accelerate interactive data exploration and machine learning work.

youtu.be/jGMf3eCXZZA

We made matplotlib interactive. You're welcome. Our favourite plotting library may just be matplotlib again now that we made it interactive with a puck! It may sounds strange at first, but pucks turn these...

Time to celebrate matplotlib again folks!

youtu.be/5ZxczGlrkyQ

How to fix your broken Python demo This is a public announcement to add devrel professionals out there: please add the right widget to deal with keys in your demos and tutorials!To explore all...

We're going to release two new widgets this week. The first one is for all you devrel-people out there that want to host a notebook as a demo.

Trust us when we say it solves a big gripe.

youtu.be/kQlCU4nAjVE