Orbit-4B (v0.1) Overview

Orbit-4B (v0.1) is a 4-billion parameter expert open search agent developed by orbit-ai, built upon the Qwen3-4B base model. It has been fine-tuned using 165 GRPO (Generalized Reinforcement Learning with Policy Optimization) steps to proficiently integrate live web search as a tool for multi-turn question answering. This model is specifically designed for research into retrieval-augmented reasoning and RL-based tool-use training.

Key Capabilities

• Multi-turn Question Answering: Excels at answering complex questions over multiple conversational turns.
• Live Web Search Integration: Utilizes a DDGS-based retriever with multiple search backends (Google, Brave, Bing, Wikipedia, Grokipedia) for real-time information gathering.
• Reinforcement Learning Fine-tuning: Trained with GRPO on a mixed dataset including Natural Questions, HotpotQA, and the ORBIT dataset, focusing on single-hop and multi-hop reasoning.
• Tool-use Protocol: Designed to issue <search> query </search> actions and process <information>…</information> observations.

Good for

• Research into RL-based tool-use training: Ideal for exploring how LLMs can effectively use external tools.
• Multi-turn retrieval-augmented reasoning: Suitable for tasks requiring iterative information retrieval and synthesis.
• Complex QA scenarios: Particularly effective for multi-hop and difficult reasoning queries where external knowledge is crucial.