OpenRSI turns AI self-improvement into an engineering loop

OpenRSI is an open-source framework that turns recursive self-improvement from a sci-fi concept into an executable engineering loop. Built specifically for machine learning engineering (MLE), it lets an AI system write, test, debug, and evolve its own programs. Best of all, it does this on consumer hardware rather than massive server farms.

What it does

OpenRSI tackles a concrete question: how can an AI systematically improve the software that builds AI? Instead of chasing unconstrained, open-ended intelligence, it anchors the problem inside MLE, where code either runs or breaks and benchmark scores provide unambiguous ground truth.

At the center of the system is Frontis-MA1, a post-trained model designed to act as an evolution engine across four core program operators: Draft (write from scratch), Improve (refine code using runtime logs), Debug (repair errors), and Crossover (splice two working scripts together).

The key design choice is shared action space. The exact same operators used during inference-time tree search are baked into the model during post-training via execution-grounded SFT and online RL. Instead of dumping entire conversation histories into context, the agent queries bounded execution evidence per operator, reducing token usage by 41.7% while increasing effective discoveries per million tokens by 84.3% compared to baseline setups like AIRA-Evo.

Specs and performance

The team released Frontis-MA1 weights in 35B and 30B sizes, offering standard BF16 checkpoints alongside GGUF derivatives for local inference.

Its headline benchmarks were achieved under strict resource caps: a single Nvidia RTX 4090 restricted to 12GB VRAM, with a 12-hour compute budget per task.

On MLE-Bench Lite, swapping the base model for Frontis-MA1 35B within the OpenMLE-Evo harness pushed the Medal Average from 39.39% to 60.61%. Turning on OpenMLE-Evo-Max—which enables asynchronous multi-GPU search and benchmark-agnostic priors—pushed that score to 71.21%. That edges past GPT-5.5 + Codex and nears frontier marks set by GPT-5.6 Sol and the 2.8T Kimi K3.

On the held-out NatureBench Lite transfer suite, the components separated cleanly: holding the framework fixed and upgrading the model raised the Match-SOTA score from 50% to 70%. Holding the base model fixed and adding the OpenMLE-Evo search framework moved it from 20% to 50%.

The authors are careful to note that these are model-plus-harness system metrics, not standalone zero-shot generations, and they explicitly disclaim solving general RSI.

How to run it

Everything is available in the FrontisAI/OpenRSI repository on GitHub, including the OpenMLE-Gym task pipeline, OpenMLE-ERL training scripts, OpenMLE-Evo search harness, and the self-hosted OpenMLE Sandbox execution backend.

You can clone the repo, set up a Python 3.12 environment, and launch a local single-task smoke test on NatureBench using their provided script. It hooks directly into local SGLang endpoints or remote OpenAI-compatible APIs, keeping all execution and evaluation local.

Why it matters

If you build autonomous coding agents or study post-training search loops, OpenRSI is a must-read release. While startups raise hundreds of millions on abstract RSI hype, this project shows how to ground self-improvement in measurable, repeatable software iterations using fully open weights and datasets.

We think the team's greatest win is compute efficiency. By replacing naive context dumping with bounded operator queries, a 35B model on a single 12GB GPU managed to go toe-to-toe with frontier models in empirical engineering. The bottleneck, however, remains sandbox feedback: this recipe thrives wherever code produces deterministically measurable results, but extending it outside verifiable benchmark sandboxes remains an open challenge.