Newsletter 25

You’re reading the 25th marimo newsletter.

We’re excited to share some big news: molab, our free cloud-hosted marimo notebook workspace, now has GPUs!

molab on CoreWeave Cloud

We’ve upgraded molab to run on CoreWeave Cloud. This new and improved molab is now in public preview; it is the default experience for all molab users. Users now get 4 CPUs, 32 GB of RAM, opt-in GPUs, faster startup times, and 12 hour sessions.

Our public preview includes many other features. Connect AI agents like Claude Code, Codex, or OpenCode to running molab notebooks using marimo pair. Use free and fast AI assistance in the molab editor via open-source models like Kimi-K2.6 hosted on Weights and Biases Inference. And share notebooks seamlessly — mirror from GitHub, share as interactive slides or apps, embed in static sites, and fork from our gallery. Notebooks on molab remain public but not discoverable, similar to GitHub Gists.

For the community. We’re providing molab for free, GPUs and all. It will remain free as long as our resources are not oversubscribed, though we may tweak resource parameters in the public preview as we adapt to load. We can’t wait to see what you build!

Get started. Get started with notebooks from our partners:

• Finetune open-source models with Unsloth.
• Learn how to train world models with a notebook by Machine Learning at Berkeley President Tejas Prabhune.
• Tackle demanding problems in the physical sciences, like in this notebook by Axiomatic.

Read our announcement blog for the full details.

New features in marimo open source

Since our last newsletter, we’ve shipped many new features. Here are some highlights.

Pair more smoothly with agents. The “Pair with an agent” modal now has a dedicated Prompt tab, making it easier to copy the boilerplate you need to bring agents like Claude Code, Codex, or OpenCode onto a notebook (#9568). For code-mode agent edits, the kernel now enforces read-before-write protection on cells, preventing surprise overwrites (#9585).

Faster, more capable WebAssembly notebooks. WebAssembly notebooks now render their last session snapshot while Pyodide is still loading (#9502), so readers see content immediately instead of a blank page. DuckDB queries over HTTP now work inside WebAssembly (#9480), and marimo export html-wasm --execute pre-runs cells so static previews don’t start blank (#9437). Exports also run in a sandboxed environment for consistent results (#9519).

Better tables. Tables gained a dedicated UI for date, datetime, and time filters (#9615) and an expanded set of column filter operators with a friendlier pill editor (#9597).

Editor quality of life. The dependency minimap now labels cells by index for easier navigation (#9633), and the long-requested cut-cells command finally landed (#8019).

Customize Mermaid diagram themes. Mermaid diagrams gained full theme customization, contributed by Will Dean (#9478). Thanks Will!

New AI models. marimo now ships out-of-the-box support for gpt-5.5 (#9488) and the latest W&B inference models (#9465).

Better observability for self-hosted deployments. marimo now exports traces via OTLP with W3C trace-context propagation (#9414), and the kernel classifies and notifies on exit reasons for easier debugging (#9500).

Community roundup

We’re at 20k+ GitHub stars, 280+ contributors, and 3.9k+ members in Discord — join the conversation!

Here are some mentions of marimo in the wild that we found particularly inspiring.

AI research benchmarks. Mario Edoardo Pandolfo and the SPAICOM Lab at Sapienza University of Rome released SEMASIA, a 5TB+ benchmark of embeddings from ~1,700 pretrained vision models across 8 datasets for latent space alignment research.

Interactive data visualization. Fritz Lekschas released jupyter-scatter v1.0, enabling first-class marimo integration for interactive 2D scatter plots. He also launched dtour, a research-grade high-dimensional data navigation tool featuring overview gallery, geodesic tours, and grand tour modes.

Single-cell genomics. Dr. Badran M.E. released scanpy-done-right, a fully modern PBMC 3k single-cell pipeline that swaps Jupyter for marimo, pandas for polars, and seaborn for plotnine.

Scientific R&D framework. Julien Sigüenza of SUFFISCIENS built nuRemics, an open-source Python framework for engineering R&D. He also demoed the full pipeline with transient laminar flow through a 90-degree cylindrical bend as the case study.

Developmental biology. Researchers at Institut Pasteur and Université de Toulouse published a STAR Protocols paper in Cell on olfactory placode morphogenesis in zebrafish, with marimo as the core interface for running, calibrating, and visualizing the agent-based simulation.

Geometric machine learning. Thejas Nagesh Gowda presented “Distance Without Flatness: Metric-Driven Clustering on Spheres, Tori and Hyperbolic Spaces,” letting attendees explore K-Means, K-Medoids, and GMM across spherical, toroidal, and hyperbolic geometries.

Computational design. Luca Florio ran Rhino 3D from inside a marimo notebook via RhinoInside, the first public demo of parametric CAD generation wired directly into a reactive notebook environment.

Energy open data. Mayk Thewessen built Marktstammdatenplotter, a live interactive plotter for Germany’s wind and solar energy market open data, deployed as a public marimo web app.

PyData Boston x Cursor hackathon. Two marimo projects stood out: Harry Joshi tied for 1st place with an Infrastructure Reliability Detector that surfaced failure events from millions of 2026 transit trips, and Calvin Van and team built Matchachusetts, a deployed Boston tourism guide for FIFA World Cup visitors.