VIDRAFT Open-Sources Darwin-180B-RSI-R3 MoE Model with Self-Improvement

On November 7, 2026, DEV Community's AI channel published an article titled "VIDRAFT Releases the Open-Source MoE Model Darwin-180B-RSI-R3", introducing Darwin-180B-RSI-R3 from VIDRAFT. According to the piece, Darwin-180B-RSI-R3 is a 180B-parameter open-weight MoE model that uses Recursive Self-Improvement (RSI) to autonomously strengthen its own reasoning weights without any new human-labeled data. In SuperCLUE's evaluation for November 2026, it ranked second globally on math and science reasoning, outperforming models from Google, DeepSeek, and xAI. The same report states that the model weights have been made publicly available on Hugging Face, with a permissive open-weight license attached. At this stage, the launch of Darwin-180B-RSI-R3 is documented only through this single report, so the technical claims and the evaluation ranking rest on that one account. No additional sources cover the model's follow-up status, and the material does not describe how the RSI mechanism is implemented in practice.