AI BriefingBAAI Releases AREX-2: A 27B Long-Horizon Agent Model Based on Qwen3.8
Key point
The model features a 262,144-token context length and learns to improve solutions through iterative test-time reflection.
Details
BAAI has released AREX-2, a 27B-parameter long-horizon agent model based on the Qwen3.8 architecture. The model is designed to improve solutions over multiple test-time rounds by proposing, measuring, reflecting, and revising its outputs.
Training and Capabilities
AREX-2 is trained on machine-learning and algorithmic-programming tasks that provide verifiable feedback, alongside existing AREX deep-research data. This training enables cross-domain performance, where skills learned in coding tasks transfer to deep research without requiring new search trajectories. Key capabilities include:
- Self-improvement: Converts extra test-time rounds into refined solutions.
- Feedback-driven reflection: Analyzes scores, logs, errors, and timings to determine next steps.
- Long-horizon reasoning: Maintains productive iteration as task complexity increases.
Technical Specifications
- Architecture: Dense Qwen3.8-compatible multimodal model
- Parameters: 27B
- Context length: 262,144 tokens
Quantized versions are available via Hugging Face (e.g., mradermacher/AREX-2-GGUF).
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
GLM-5.3-Flash Matches Specialized Decision Models in Zero-Shot Classification via Logit ExtractionWhiteboard Launches Open-Source VS Code Fork for Human-AI Collaborative Software Design