# AREX-2

> BAAI's 27B open-weight agent model, released September 29, 2026: fine-tuned from Qwen3.8-27B to improve a solution over many test-time rounds of propose, measure, reflect and revise, under Apache 2.0.

- Maker: Beijing Academy of Artificial Intelligence (BAAI)
- Line: AREX
- Canonical page: https://ai-tldr.dev/models/arex-2/
- Updated: 2026-09-29

## In short

AREX-2 is BAAI's 27B open-weight agent model, released September 29, 2026: fine-tuned from Qwen3.8-27B to improve a solution over many test-time rounds of propose, measure, reflect and revise, under Apache 2.0. It has a 262,144-token context window. Its weights can be downloaded under the Apache-2.0 licence.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 262,144 tokens | [1] |
| Licence | Apache-2.0 · open weights | [1] |
| Released | 29 September 2026 | [2] |
| Parameters | 27B (dense) | [1] |

## Sources

1. [Hugging Face — BAAI/AREX-2](https://huggingface.co/BAAI/AREX-2)
2. [GitHub — VectorSpaceLab/AREX-2 (evaluation code)](https://github.com/VectorSpaceLab/AREX-2)

## Overview

AREX-2 is a 27-billion-parameter long-horizon agent model from the Beijing Academy of Artificial Intelligence (BAAI), published on Hugging Face on September 29, 2026 under the Apache 2.0 license together with the paper "AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks". It is a dense, Qwen3.8-compatible multimodal model built on Qwen3.8-27B, with a 262,144-token context window.

The model is trained to turn extra test-time rounds into a better answer: it proposes a solution, measures it, reflects on the feedback and revises. BAAI's paper treats this as two connected abilities — reflection, which reads scores, logs, errors and timings to decide what to change, and long-horizon execution, which keeps the improvement loop useful as the task budget grows.

BAAI trained AREX-2 on machine-learning-engineering and algorithmic-programming tasks where progress can be checked directly, together with the existing AREX deep-research data. According to the model card, the self-improvement behaviour learned on coding and ML tasks transfers to deep research without adding new search trajectories.

In BAAI's model card AREX-2 scores 70.7 on the 188-task Frontier-CS Agent Track and 81.8 on MLE-Lite (Any Medal, averaged over three seeds), and 84.0 on BrowseComp, 52.6 on HLE (text-only subset), 92.2 on GAIA and 93.8 on DeepSearchQA. The same tables list GPT-5.6 Sol at 76.4 on Frontier-CS and Kimi-K3 at 91.2 on BrowseComp, so the 27B model does not lead every column.

Alongside the weights, BAAI publishes the evaluation runners and dataset definitions for the six reported tracks in the VectorSpaceLab/AREX-2 GitHub repository. The model card positions AREX-2 for research on long-horizon agents, iterative problem solving, ML engineering, algorithmic coding and tool-augmented deep research.

## Benchmarks

- DeepSearchQA: 93.8% ([source](https://huggingface.co/BAAI/AREX-2))
- GAIA: 92.2% ([source](https://huggingface.co/BAAI/AREX-2))
- BrowseComp: 84% ([source](https://huggingface.co/BAAI/AREX-2))
- MLE-Lite (Any Medal): 81.8% ([source](https://huggingface.co/BAAI/AREX-2))
- Frontier-CS (Agent Track): 70.7 ([source](https://huggingface.co/BAAI/AREX-2))
- HLE (text-only): 52.6% ([source](https://huggingface.co/BAAI/AREX-2))

## Strengths

- Trained to refine its own solution over many rounds, reading scores, logs, errors and timings to decide what to change
- 81.8 on MLE-Lite in BAAI's table, above every closed and open model listed there
- 84.0 on BrowseComp, 92.2 on GAIA and 93.8 on DeepSearchQA at 27B parameters
- Apache 2.0 weights with a 262,144-token context and open evaluation code

## Best for

- Reach for it for self-hosted agent loops that iterate on a measurable target — ML experiments, algorithmic coding against a judge, or benchmarked optimisation.
- Reach for it for tool-augmented deep research where a 27B open model fits your hardware.
- Look elsewhere for a general chat assistant: BAAI frames AREX-2 as a research model for long-horizon agents, not a tuned consumer assistant.

## AREX versions

- AREX-2 (29 September 2026): 27B dense agent model trained for long-horizon self-improvement, fine-tuned from Qwen3.8-27B. — this page

## FAQ

### What is AREX-2?

AREX-2 is a 27B-parameter open-weight agent model from the Beijing Academy of Artificial Intelligence (BAAI), published on Hugging Face on September 29, 2026. It is fine-tuned from Qwen3.8-27B to improve a solution over many test-time rounds: propose, measure, reflect and revise.

### What license is AREX-2 released under?

Apache 2.0. BAAI's model card also asks users to follow the terms of the Qwen base model and of any downstream data or tools.

### How does AREX-2 score on agent benchmarks?

In BAAI's model card it scores 70.7 on Frontier-CS (Agent Track), 81.8 on MLE-Lite, 84.0 on BrowseComp, 52.6 on HLE (text-only subset), 92.2 on GAIA and 93.8 on DeepSearchQA. In the same tables GPT-5.6 Sol leads Frontier-CS at 76.4 and Kimi-K3 leads BrowseComp at 91.2.

### How do I run AREX-2?

The model card loads it with Hugging Face Transformers (AutoModelForMultimodalLM and AutoProcessor with the model id BAAI/AREX-2, in bfloat16) and asks for a recent Transformers release with Qwen3.8 support. The weights are BF16 safetensors.

## Links

- [Hugging Face — BAAI/AREX-2](https://huggingface.co/BAAI/AREX-2)
- [arXiv — AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks](https://arxiv.org/abs/2609.38288)
- [GitHub — VectorSpaceLab/AREX-2 (evaluation code)](https://github.com/VectorSpaceLab/AREX-2)
- [AREX-2 project site](https://vectorspacelab.github.io/AREX-2/)
- [Beijing Academy of Artificial Intelligence](https://www.baai.ac.cn)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/arex-2/