BAAI releases AREX-2, an open model for agents that improve themselves on long tasks

The Beijing Academy of Artificial Intelligence (BAAI) released AREX-2 on Hugging Face on September 29, 2026. The 27-billion-parameter model is designed as an agent that improves a solution over several rounds on long tasks: it proposes a solution, measures the result, reads scores, logs, errors and timings to decide what to change, and revises. It is offered as a model you can run on your own servers.

A self-improving agent model

According to the model page, AREX-2 was trained on machine-learning engineering and algorithmic programming tasks with verifiable feedback, together with earlier AREX deep-research data. BAAI says the self-improvement behavior learned from these tasks carries over to deep research without adding new search trajectories. The model is trained to keep improving the solution productively as the number of rounds grows. Details of the method are in BAAI's arXiv paper, and the evaluation scripts are in the GitHub repository.

Technical details

AREX-2 is built on Alibaba's Qwen3.8-27B and is a dense model, meaning all parameters are used at every step. It uses a multimodal architecture, so it can take images as well as text. The context window is 262,144 tokens. It is licensed under Apache 2.0 and commercial use is allowed; BAAI asks users to also follow the terms of the Qwen base model.

At full precision (BF16) the model takes about 55 GB. The card gives sample code for Hugging Face transformers. Since the model uses the Qwen3.8 architecture, it is expected to run on serving tools that support it, such as vLLM, though the card has no separate instructions for this. BAAI has not released a quantized version; once a 4-bit version is available, the model is expected to fit on a single 24 GB GPU.

In BAAI's tables, AREX-2 stands out among models under 40 billion parameters on deep-research tests such as BrowseComp, HLE and DeepSearchQA, while trailing some similarly sized models on GAIA. On coding and machine-learning tests (Frontier-CS and MLE-Lite) it appears ahead of much larger open models in the table; on MLE-Lite it also tops the closed models listed, while on Frontier-CS it trails GPT-5.6 Sol and Claude Opus 4.8. For companies that want to run research and coding agents on their own servers without sending data outside, it is an option sized to run on a single strong server.

All results were produced with the methods in BAAI's own paper, and on the deep-research tests an external judge model scores the answers. The model card describes AREX-2's intended use as research on long-horizon agents, machine-learning engineering and tool-augmented deep research, and users need to set up the search and tool infrastructure themselves. The model is newly released and there is no community feedback yet. To see how it does on different tasks, it is worth following evaluations from independent researchers.

Sources: Hugging Face: AREX-2 model card, GitHub: VectorSpaceLab/AREX-2, arXiv: AREX paper.