AREX-2
BAAI's 27B agent built on Qwen3.8 that improves its own answer over many rounds. 81.8 on MLE-Bench Lite, 84.0 on BrowseComp.
In short. AREX-2 is an agent model that does not stop at its first answer. It proposes a solution, checks the score or the error log, works out what went wrong and tries again, for as many rounds as you allow. It is free for commercial use under Apache 2.0, and a 27B dense model that fits on one large GPU.
Summary
AREX-2 is a 27-billion-parameter agent model from the Beijing Academy of Artificial Intelligence (BAAI), published on 29 September 2026 under the Apache 2.0 license. It is built on Qwen3.8-27B and trained for one skill: improving its own answer over many rounds. In each round it proposes a solution, measures the result using scores, logs, errors and timings, reflects on what went wrong, and revises.
To teach this, BAAI's team synthesized long improvement trajectories from machine-learning engineering and algorithmic programming tasks, two areas where a program can check whether a revision really is better. It combined them with the data from the earlier AREX deep-research models and added no new search trajectories, yet the card says the behavior carried over to web research. The model keeps the base model's image input and 262,144-token context, and the paper says it keeps improving as its budget of rounds grows.
BAAI's tables report 70.7 on the 188-task agent track of Frontier-CS and 81.8 on MLE-Bench Lite (Any Medal, averaged over three seeds), the highest MLE-Bench Lite score in its table, where the next best is 73.7. For deep research it reports 84.0 on BrowseComp, 52.6 on the text-only subset of Humanity's Last Exam, 92.2 on GAIA and 93.8 on DeepSearchQA. Among models of 40B or less in the same table it leads on BrowseComp, Humanity's Last Exam and DeepSearchQA, but trails two others on GAIA (97.1 and 96.0). On Frontier-CS, closed models such as GPT-5.6 Sol (76.4) and Claude Opus 4.8 (74.5) are still ahead. All of the results are BAAI's own.
Good for
- Research on agents that refine a solution over many rounds, such as machine-learning experiments where each run returns a score
- Algorithmic programming and machine-learning engineering agents that can read scores, logs, errors and timings
- Deep-research agents, once you supply your own search and browsing tools
- A dense 27B model that fits on one 80 GB GPU in BF16 (about 55 GB of weights)
Watch out for
- The card describes it as a research model, and the scores follow protocols from the paper, so a plain chat call will not reproduce them
- It trails closed frontier models on Frontier-CS and two smaller models on GAIA in BAAI's own table
- The card asks you to also follow the terms and notices of the Qwen base model and of any data or tools involved
- All benchmark numbers are self-reported by BAAI
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the AREX-2 (27B) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| AREX-2 (27B) BF16 | 27.4B | 63 GB | 32 GB | 19 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is AREX-2?
- AREX-2 is an agent model that does not stop at its first answer. It proposes a solution, checks the score or the error log, works out what went wrong and tries again, for as many rounds as you allow. It is free for commercial use under Apache 2.0, and a 27B dense model that fits on one large GPU.
- Can I use AREX-2 commercially?
- Yes. AREX-2 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is AREX-2 open source?
- AREX-2 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run AREX-2 on my own computer?
- Yes, on a good graphics card. The smallest variant, AREX-2 (27B), needs about 19 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was AREX-2 released?
- AREX-2 was released by BAAI on September 29, 2026.
- How long a context window does AREX-2 have?
- AREX-2 supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- AREX-2 model card (Hugging Face) huggingface.co
- AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks (arXiv) arxiv.org
- AREX-2 repository (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Nanbeige4.3-33B Nanbeige LLM Lab | 33.1B | 512K | Apache 2.0 (Permissive) | 23 GB Gaming GPU | |
| Kimi K3 Moonshot AI | 2.8T 104B active | 1M | Kimi K3 License (Conditional) | 1.72 TB Multi-node cluster | |
| GLM-5.3-Flash Z.ai (Zhipu AI) | 320B 18B active | 1M | MIT (Permissive) | 221 GB Big-memory desktop |