Sam Witteveen · 2026-11-06 · notable
Sam Witteveen — 'AREX-2 - The Harness in the Model'
In short
Sam Witteveen explains AREX-2, BAAI's 27B open-weight model fine-tuned from Qwen3.8-27B that runs the try, check and retry loop itself instead of leaving it to the agent harness.
Sam Witteveen · 2026-11-06 · 4 sources checked
Sam Witteveen looks at AREX-2, a model trained to do the agent loop that harnesses usually do.
What is it?
'AREX-2 - The Harness in the Model' went up on Sam Witteveen's channel on 6 November 2026. It covers AREX-2 from the Beijing Academy of Artificial Intelligence (BAAI), a 27B Apache 2.0 model built on Qwen3.8-27B with a 262,144-token context window.
How does it work?
The video moves through who built the model, the core idea of putting reflection and long-horizon retries inside the model, the three-step data pipeline, benchmarks, and skills versus training. Witteveen also asks whether this counts as recursive self-improvement and runs a hands-on test.
Why does it matter?
Agent quality today depends heavily on the harness around a model. AREX-2 tries to train that behaviour into open weights, and BAAI's model card lists 84.0 on BrowseComp and 92.2 on GAIA, so the video helps viewers judge the idea before trying the weights.
Who is it for?
developers building agents on open-weight models