LiteResearcher-4B is a 4-billion parameter deep research agent developed by wanlilll, built on the Qwen3ForCausalLM architecture with a 262,144-token context window. It is trained using scalable agentic reinforcement learning, enabling it to perform web-based research tasks. This model excels at complex information retrieval and synthesis, matching or exceeding larger proprietary models like Claude-4.5-Sonnet on benchmarks such as GAIA-Text and Xbench-DS, despite its significantly smaller size.
Loading preview...
LiteResearcher-4B: A Compact Deep Research Agent
LiteResearcher-4B is a 4-billion parameter model developed by wanlilll, designed as a deep research agent. It leverages a Qwen3ForCausalLM architecture and boasts an extensive 262,144-token maximum context length. A key differentiator is its training via scalable agentic reinforcement learning, which allows it to operate as a ReAct-style agent, iteratively thinking, searching the web, visiting webpages for evidence, and synthesizing answers.
Key Capabilities & Performance
- Exceptional Research Prowess: Despite its small size, LiteResearcher-4B matches Claude-4.5-Sonnet on GAIA-Text (71.3%) and outperforms open-source models up to 8 times larger on benchmarks like Xbench-DS (78.0%) and WebWalkerQA (72.7%).
- Agentic Workflow: Utilizes
searchandvisittools within a ReAct framework, structuring its reasoning with<think>,<tool_call>, and<answer>tags. - Scalable Training: Employs a two-stage difficulty-aware curriculum RL with a virtual world environment, trained on a vast corpus of over 32 million webpages and 1 million domains.
Good For
- Complex Information Retrieval: Ideal for applications requiring deep web research, evidence extraction, and synthesis of information from multiple sources.
- Resource-Constrained Environments: Its small parameter count (4B) makes it suitable for scenarios where larger models are impractical, offering high performance with reduced computational overhead.
- Automated Research Tasks: Can be integrated into systems needing autonomous agents capable of navigating and understanding web content to answer intricate queries.