AREX-2 - The Harness in the Model

In this video, I go through AREX-2, a new 27B open-weights model from BAAI, fine-tuned from Qwen3.8-27B built to run the try, check and retry loop itself rather than leaving it all to your harness. Paper: arxiv.org/abs/2609.38288 HF Link: huggingface.co/BAAI/AREX-2 GitHub: github.com/VectorSpaceLab/AREX-2 2nd Channel: youtube.com/channel/UCUF7WtR4JpVbQqcQGrLeADw Twitter: x.com/Sam_Witteveen 🕵️ Interested in building LLM Agents? Fill out the form below Building LLM Agents Form: drp.li/dIMes 👨‍💻Github: github.com/samwit/llm-tutorials ⏱️Time Stamps: 0:00 Intro 0:52 AREX-2: a 27B Qwen fine-tune 1:38 Who built it 2:24 Core idea 2:44 Reflection and long horizon 5:41 The three-step data pipeline 7:08 Benchmarks 8:35 Skills vs training 11:30 Is this really RSI? 12:32 Testing it: Round Lab 14:42 Takeaways