BAAI/AREX-Turbo

AREX-Turbo: A Compact Deep Research Agent

AREX-Turbo is a 4 billion parameter dense model developed by the Beijing Academy of Artificial Intelligence (BAAI), built upon the Qwen3.5-4B backbone. It features an extensive 262,144 token context length and is designed as a compact member of the AREX family of deep research agents. The model is engineered for long-horizon tasks that involve searching across diverse sources, synthesizing information, verifying multiple constraints, and dynamically revising research plans based on incomplete evidence.

Key Capabilities

• Recursive Self-Improvement: Employs an inner research loop for evidence integration and an outer self-improvement loop to evaluate and refine research trajectories.
• Verification-Guided Research: Transforms unresolved constraints into targeted follow-up searches, enhancing research efficiency.
• Autonomous Context Management: Automatically updates its research state with verified findings, rejected candidates, and evolving research plans.
• Long-Horizon Tool Use: Supports multi-round interactions for searching, browsing, integrating evidence, and constructing answers.
• Cost-Effective Deployment: Provides AREX's advanced research capabilities in a smaller 4B model, optimizing for lower serving costs and faster iteration.

When to Use AREX-Turbo

AREX-Turbo is particularly well-suited for research-agent applications where:

• Long-horizon information seeking and evidence aggregation are critical.
• Multi-constraint verification and tool-augmented reasoning are required.
• Serving cost, latency, or iteration speed are important considerations for deployment.