OpenRSI Open-Sources 35B Model: Beats GPT-5.5 on Single RTX 4090
aigclink · x · 2026-08-09
Tsinghua University and Frontis open-sourced OpenRSI, a project turning Recursive Self-Improvement (RSI) into an executable engineering problem.
- Key Results: Their 35B model, running on a single RTX 4090 (12GB VRAM limit) with 12 hours per task, surpasses GPT-5.5+Codex on the MLE benchmark and approaches GPT-5.6 Sol and the 2.8T Kimi K3.
- Architecture: Defines four atomic operators (Draft / Improve / Debug / Crossover) with a shared interface for training and inference, making the model the mutation engine of its own evolution framework.
- Experiments: Clean ablation studies separate model vs. framework contributions. Compared to AIRA-Evo, total token consumption dropped 41.7%, but effective discoveries per million tokens increased by 84.3%.
The authors note this is currently a system score of model + harness, explicitly stating general RSI is not yet achieved, focusing more on meta-self-evolution.
More from coding & agent
- ComfyUI-SigmaSync-LoRA: Schedule LoRA Strength Against Sampler SIGMAS — Capitan01R- · 2026-08-09
- Prime Agent: Open-Source Self-Improving RLM Agent for Coding, 2,293 Stars in 24h — Roger_M_Taylor · 2026-08-09
- NVIDIA Open-Sources NOOA: An Agent Framework Treating Agents as Python Classes — AccBalanced · 2026-08-09
- Lonkero: Modular Rust Web Security Scanner with ML Integration — tom_doerr · 2026-08-09
- Optimizing Video Generation on RTX 3090: Low-Res Generation and Upscaling — cocktailpeanut · 2026-08-09
- Reproducing Papers to Code: RepLLM Multi-Agent Framework — jiqizhixin · 2026-08-09