FULL STORY

OpenMLE System and Frontis-MA1 Model Released

Tsinghua researchers released the open-source OpenMLE system and the 35B Frontis-MA1 model to explore AI recursive self-improvement.

2026-07-31 ~ 2026-08-04 · 3 episodes · 7 posts

Episode 1 · FrontisAI Open-Sources OpenMLE System and 35B Model (2026-07-31, 2 posts)

FrontisAI has open-sourced OpenMLE, a full-stack AI4AI system for recursive self-improvement, alongside the Frontis-MA1 35B model. This system leverages evolutionary agents to achieve a 71% win rate on MLE-Bench Lite using a single RTX 4090, approaching GPT-5.6 performance.

Episode 2 · OpenMLE: An Open-Source Stack for Recursive AI Self-Improvement (2026-08-01, 3 posts)

Researchers released OpenMLE, an open-source full-stack system providing an executable testbed for recursive self-improvement. Alongside Frontis-MA1, an AI4AI agent designed for machine learning engineering, the project aims to optimize the AI building process and can run efficiently on a single RTX 4090.

Episode 3 · Tsinghua Releases Frontis-MA1 for Recursive Self-Improvement (2026-08-03, 2 posts)

Tsinghua University introduced Frontis-MA1, a 35B-parameter model designed for recursive self-improvement in machine learning engineering. By integrating evolutionary algorithms, the model outperforms GPT-5.5 and unlocks greater potential beyond traditional hill-climbing approaches.