II-Search-CIR-4B: Code-Integrated Reasoning for Enhanced Search

II-Search-CIR-4B is a 4-billion parameter model from Intelligent Internet, built upon the Qwen3-4B architecture. It significantly advances tool interaction through its unique Code-Integrated Reasoning (CIR) methodology. Unlike traditional tool-calling paradigms, CIR enables the model to generate and execute Python code blocks, allowing for programmatic interaction with external resources.

Key Capabilities:

• Code-Integrated Reasoning (CIR): The model generates code blocks (e.g., web_search, web_visit) to interact with external tools, process information, and reason programmatically.
• Enhanced Information Seeking: Specifically fine-tuned to excel in tasks requiring external information retrieval and processing.
• Improved Performance: Outperforms its base model (Qwen3-4B) and other small-sized search-specialized models like Jan-4B and WebSailor-3B on benchmarks such as Google Frames and Seal_0.

Training Methodology:

The model underwent a two-stage training process:

1. SFT Fine-tuning: Initial Supervised Fine-Tuning on a curated dataset to efficiently produce the required code format.
2. DAPO Optimization: Further optimized using DAPO (Direct Advantage Policy Optimization) on a hard-reasoning dataset to boost performance.

Good for:

• Applications requiring advanced web search and information synthesis.
• Developing agents that need to programmatically interact with external APIs or data sources.
• Tasks where reasoning over retrieved information is critical, beyond simple retrieval.

For more details on the training methodology and datasets, refer to the II-Search-4B blog post and the released datasets: II-Search-CIR-SFT and II-Search-RL.