SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Tongyi DeepResearch - Alibaba's Open-Source Deep Research Agent

Author: AI小集 · Original source: AI工具集

What is Tongyi DeepResearch

Tongyi DeepResearch is an open-source deep research agent launched by Alibaba, designed specifically for long-horizon, deep information retrieval tasks. It has 30 billion parameters, activates 3 billion parameters each time, and supports ReAct mode and Heavy Mode, the latter enhancing complex reasoning capabilities through an iterative research paradigm (IterResearch). The agent adopts a full-process synthetic data solution, generating high-quality datasets without manual intervention, breaking through the upper limits of agent capabilities. The training pipeline covers Agentic Continual Pre-Training (Agentic CPT), Supervised Fine-Tuning (SFT), and Reinforcement Learning (RL), forming a complete end-to-end training chain. Tongyi DeepResearch has empowered multiple applications within Alibaba, such as Amap's AI-native travel Agent and the legal domain's "Tongyi Farui”。

Tongyi DeepResearch

Main Features of Tongyi DeepResearch

  • Long-horizon deep information retrieval: Designed for complex, long-horizon information retrieval tasks, capable of handling multi-step reasoning and planning, suitable for scenarios such as academic research, market analysis, and policy making.
  • Multi-mode reasoning support: Supports ReAct mode and Heavy Mode. ReAct mode strictly follows the "thought-action-observation" loop, suitable for evaluating the model's core capabilities; Heavy Mode enhances complex reasoning capabilities through an iterative research paradigm (IterResearch).
  • Full-process synthetic data generation: Adopts a self-developed full-process synthetic data solution, generating high-quality datasets without manual intervention, breaking through the upper limits of agent capabilities, and supporting a complete training chain from pre-training to fine-tuning to reinforcement learning.
  • End-to-end reinforcement learning: Through customized reinforcement learning algorithms (such as Group Relative Policy Optimization, GRPO), ensures that the agent's behavior remains consistent with high-level goals, enhancing the model's adaptability and stability in dynamic environments.
  • Practical application empowerment: Has been successfully applied to multiple scenarios within Alibaba, such as Amap's AI-native travel Agent and the legal domain's "Tongyi Farui", demonstrating strong practicality and value.
  • Open-source co-construction: The project is fully open-source, providing complete code, models, and data, encouraging developers to participate in co-construction, and promoting the development and innovation of deep research agents.

Technical Principles of Tongyi DeepResearch

  • Full-process synthetic data solution: Without manual intervention, automatically generates high-quality datasets, supporting a complete training chain from pre-training to fine-tuning to reinforcement learning, breaking through the upper limits of agent capabilities.
  • Iterative research paradigm (IterResearch): Decomposes complex tasks into multiple research rounds, each round dynamically reconstructing a concise workspace, and through the "thought-synthesis-action" process, enhances complex reasoning capabilities and decision quality.
  • End-to-end reinforcement learning: Adopts customized reinforcement learning algorithms, such as Group Relative Policy Optimization (GRPO), ensuring that the learning signal precisely matches the model's current capabilities, enhancing the model's adaptability and stability in dynamic environments.
  • Large-scale continual pre-training: Utilizes continuously updated knowledge documents, crawler data, knowledge graphs, etc., to build open-world knowledge memory, generate multi-style (question, answer) pairs, and continuously expand model capabilities.
  • Automated data management: Optimize data in real time under the guidance of training dynamics, and dynamically adjust the training set through fully automated data synthesis and data funnels, ensuring training stability and performance improvement.
  • Stable and efficient tool sandbox: Develop a unified sandbox environment to handle concurrency and failures, ensure the stability and reliability of tool calls, and provide agents with a fast and robust interaction environment.

Tongyi DeepResearch project address

  • Project official website：https://tongyi-agent.github.io/blog/introducing-tongyi-deep-research/
  • Github repository：https://github.com/Alibaba-NLP/DeepResearch
  • HuggingFace model library：https://huggingface.co/Alibaba-NLP/Tongyi-DeepResearch-30B-A3B

Family members of Tongyi DeepResearch

  • WebWalker: Focuses on web traversal tasks, used to evaluate language model performance in web navigation.
  • WebDancer: Committed to achieving autonomous information-seeking capabilities, promoting agent autonomy in information retrieval.
  • WebSailor: Used to navigate complex web environments and enhance agents' superhuman-level reasoning capabilities.
  • WebShaper: Through the formalization of information seeking, realize the synthesis of agent data and improve data quality and model performance.
  • WebWatcher: Explore new frontiers of visual-language agents, combining visual and language capabilities for deep research.
  • WebResearcher: Unlock the unbounded reasoning capabilities of long-horizon agents and improve their performance in complex tasks.
  • ReSum: Unlock long-horizon search intelligence through context summarization and optimize agents' information management capabilities.
  • WebWeaver: Use dynamic outlines to structure web-scale evidence and support open-ended deep research.
  • WebSailor-V2: Narrow the gap with proprietary agents through synthetic data and scalable reinforcement learning.

Application scenarios of Tongyi DeepResearch

  • Academic research: Can quickly organize literature reviews, helping scholars efficiently complete complex academic research tasks and improve research efficiency.
  • Market analysis: Provides enterprises with competitor analysis, industry trend reports, etc., helping enterprises formulate precise market strategies.
  • Legal research: In the legal field, such as the "Tongyi Farui" application, automatically retrieve legal provisions, similar cases, and judicial documents, conduct in-depth inductive analysis, and provide legal practitioners with a powerful productivity tool.
  • Travel planning: In cooperation with Amap, launch an AI-native travel Agent, combining real-time data to provide users with precise travel suggestions and planning.
  • Complex information retrieval: Suitable for complex information retrieval tasks requiring multi-step reasoning and planning, such as cross-domain research, policy making, etc., helping users quickly obtain and integrate information.