Awesome AI Agents › AI Agent Development

Alibaba-NLP/DeepResearch

⭐ 19933 Python added to this list on 2025-10-04 repository created 2025-01-09

Tongyi DeepResearch is an advanced open-source large language model developed by Alibaba's Tongyi Lab, designed specifically for deep, long-horizon information-seeking tasks. The model features 30.5 billion parameters, with an efficient activation of only 3.3 billion per token, enabling powerful yet resource-conscious performance. It excels in agentic search benchmarks such as Humanity's Last Exam, BrowserComp, WebWalkerQA, and others, demonstrating state-of-the-art capabilities in complex information retrieval and reasoning. The project builds upon the previous WebAgent initiative and introduces several innovative features. It includes a fully automated synthetic data generation pipeline that supports agentic pre-training, supervised fine-tuning, and reinforcement learning, enabling scalable and high-quality training data creation. The model undergoes large-scale continual pre-training on diverse agentic interaction data to enhance its reasoning abilities and maintain up-to-date knowledge. A key highlight is the end-to-end reinforcement learning approach using a customized Group Relative Policy Optimization framework. This method incorporates token-level policy gradients, leave-one-out advantage estimation, and selective filtering of negative samples to stabilize training in dynamic environments. Tongyi DeepResearch supports two inference paradigms: ReAct, which rigorously evaluates the model's intrinsic abilities, and an IterResearch-based 'Heavy' mode that applies test-time scaling to maximize performance. The model is accessible via HuggingFace, ModelScope, and Alibaba's Bailian service, with options for online demos and local deployment for stable and production-ready use. It supports a large context length of 128K tokens, making it suitable for extensive document understanding and complex query answering. The repository provides detailed setup instructions, including environment configuration, dependency installation, and evaluation data preparation. It supports JSON and JSONL input formats for evaluation datasets. Overall, Tongyi DeepResearch represents a cutting-edge tool for researchers and developers seeking advanced AI-driven deep research and information retrieval capabilities.

https://github.com/Alibaba-NLP/DeepResearch

agentagentic-interaction-dataagentic-searchai-modelalibabaalibaba-nlpartificial-intelligencebailian-servicecontinual-pre-trainingdeep-information-seekingdeep-researchdeepresearchhuggingfaceinformation-retrievalinformation-seekingiterresearchlarge-language-modelllmlong-horizon-tasksmodelscopeopen-sourcepolicy-optimizationreact-inferencereinforcement-learningsynthetic-data-generationtongyitongyi-deepresearchweb-agentwebagent