Relm – local LLMs as base-R objects, with interpretability

R fans are losing it over a tool that lets them run chatbots without leaving their comfort zone

TLDR: relm lets R users run AI models on their own computers and get results back in familiar table-like forms, aiming squarely at researchers. Commenters are split between calling it a huge win for science workflows and roasting it as another dramatic attempt to keep R users away from Python.

A new project called relm just dropped with a very specific promise: people who live in R, the stats-heavy coding language loved in science and academia, can now run local AI models without getting dragged into Python drama. And yes, the reaction is exactly what you’d expect from the internet. One camp is cheering this like a cultural liberation moment — basically, “finally, researchers can use modern AI in the language they already know.” The biggest applause is for the tool returning plain old tables and matrices instead of weird new formats, which many commenters treated like a love letter to exhausted scientists.

But the other side showed up fast. Skeptics rolled their eyes at yet another “R is back” declaration, joking that every few years someone announces an R-enaissance like it’s a royal baby. Some questioned whether this is genuinely useful or just a very elaborate way to avoid touching Python. Others were impressed but immediately started nitpicking the setup, the version requirements, and the promise of “interpretability,” with the usual internet suspicion of anything that sounds a little too neat.

The funniest reactions were gloriously petty: jokes about statisticians refusing to leave their “spreadsheet monastery,” memes about Python users being “banished,” and lots of delight over the name R-ebirth, which commenters said sounds either visionary or like an anime reboot. Underneath the snark, though, there’s real excitement: local AI that runs on your own machine, aimed at researchers, with tools to peek inside how it behaves. Even the doubters seem to agree on one thing — this is the kind of niche release that could make a loud impact in labs.

Key Points

  • Relm is an R package with a Rust core that exposes local LLM operations as base-R functions returning standard R objects such as data.frames and matrices.
  • The first public release, relm v0.1.0, is text-only and supports local GGUF models for loading, generation, embeddings, tokenization, logits, tracing, steering, ablation, and model download.
  • The article highlights a no-Python topic-modelling workflow using llm_embed(), UMAP, and HDBSCAN, with the model naming each cluster.
  • The repository includes an R package, a Rust workspace, a vendored patched llama.cpp, numerical golden tests, and two runnable demo directories.
  • Source builds require R 4.5 or newer, a C toolchain, a Rust toolchain, and CMake 3.28 or newer, while end users can install prebuilt binaries from r-universe.

Hottest takes

"This is what happens when statisticians finally say ‘fine, we’ll build our own AI stack’" — data_druid
"An R package with a Rust core wrapping llama.cpp is the most 2020s sentence imaginable" — segfaultsoup
"Not needing Python is either the killer feature or the biggest red flag, depending on your trauma" — bayesandconfused
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.