simplex-ai-inc/LiteResearcher-4B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Apr 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

LiteResearcher-4B by simplex-ai-inc is a 4-billion parameter deep research agent built on the Qwen3ForCausalLM architecture with a 262,144 token context window. Trained using scalable agentic reinforcement learning, it functions as a ReAct agent capable of web search and information extraction. This model uniquely matches Claude-4.5-Sonnet on GAIA and outperforms open-source models up to 8x larger, making it highly efficient for complex research tasks requiring tool use.

Loading preview...

LiteResearcher-4B: A Compact Deep Research Agent

LiteResearcher-4B, developed by simplex-ai-inc, is a 4-billion parameter deep research agent that leverages scalable agentic reinforcement learning. Despite its small size, it demonstrates remarkable performance, matching proprietary models like Claude-4.5-Sonnet on the GAIA benchmark and outperforming open-source models up to 8 times larger.

Key Capabilities

  • Advanced Research Agent: Operates as a ReAct-style agent, iteratively thinking, searching the web (via Google), visiting webpages to gather evidence, and formulating answers.
  • Exceptional Benchmark Performance: Achieves 71.3% on GAIA-Text (matching Claude-4.5-Sonnet), 78.0% on Xbench-DS (surpassing Tongyi DeepSearch 30B), and 83.1% on Frames (exceeding Claude-4-Sonnet).
  • Large Context Window: Features a maximum context length of 262,144 tokens, enabling extensive information processing.
  • Efficient Architecture: Built on the Qwen3ForCausalLM (Qwen3-4B-Thinking base) architecture, optimized for agentic tasks.
  • Scalable Training: Utilizes a three-component framework including co-constructed training data from 32M+ webpages, a stable local tool environment for 73.2M tool calls, and difficulty-aware curriculum RL.

Good for

  • Complex Information Retrieval: Ideal for tasks requiring deep web research, evidence gathering, and synthesis from multiple sources.
  • Resource-Constrained Environments: Offers high performance for research tasks with a significantly smaller parameter count compared to larger models.
  • Agentic Workflow Development: Provides a robust base for building and deploying intelligent agents that interact with external tools like search engines.