In a significant leap for agentic AI systems, Alibaba's Tongyi Lab has open-sourced Tongyi DeepResearch—a large language model specifically engineered for long-horizon, information-intensive tasks. Unlike traditional dense models, this 30.5-billion-parameter architecture activates only 3.3 billion parameters per token through sparse activation pathways, dramatically improving efficiency while tackling complex research workflows requiring deep reasoning and evidence synthesis.

Architectural Innovation

The model’s sparse activation design allows it to dynamically route computations, making it exceptionally suited for extended research sessions where traditional LLMs face context degradation. This efficiency enables unprecedented performance across 10+ agentic benchmarks, including:
- Humanity's Last Exam (complex QA)
- BrowserComp (cross-lingual web interaction)
- WebWalkerQA (multi-step web traversal)
- xbench-DeepSearch (evidence-based reasoning)

Behind the Training Breakthroughs

Three pillars define Tongyi DeepResearch’s development:

  1. 🤖 Fully Automated Synthetic Data Pipeline
    Generates high-quality agent interaction data at scale—covering pre-training, supervised fine-tuning, and reinforcement learning phases—without human labeling bottlenecks.

  2. 🔄 Continual Pre-training
    Maintains model "freshness" through ongoing exposure to evolving web-scale data, preventing knowledge stagnation common in static LLMs.

  3. 🎚️ Token-Level Reinforcement Learning
    Uses a custom Group Relative Policy Optimization (GRPO) framework with:

leave_one_out_advantage_estimation()
selective_negative_sample_filtering()

This stabilizes training in non-stationary environments—critical for real-world agent deployment.

Dual Inference Modes

At runtime, developers can toggle between two paradigms:
- ReAct Mode: For evaluating core reasoning abilities using thought-action-observation loops
- IterResearch 'Heavy' Mode: Maximizes performance via test-time scaling strategies for mission-critical applications

Article illustration 1

Benchmark results across agentic tasks (Source: Tongyi Lab)

Access and Implementation

Available immediately on 🤗 HuggingFace and ModelScope, the model supports 128K-token contexts. Setup is streamlined:

conda create -n react_infer_env python=3.10.0
conda activate react_infer_env
pip install -r requirements.txt
bash run_react_infer.sh  # Customize tool APIs (web search, retrieval, etc.)

The Bigger Picture

This release anchors Alibaba’s broader "Deep Research Agent Family"—including WebWalker, WebSailor, and WebResearcher models—documented across 11 companion papers. The work signals a strategic shift toward specialized, tool-integrated LLMs that transcend chatbots, targeting domains like academic research, competitive intelligence, and forensic analysis.

As autonomous agents evolve from novelties to productivity tools, Tongyi DeepResearch establishes a compelling precedent: Efficiency needn’t be sacrificed for depth when architecture and training align with real-world use cases. The true test? Whether developers can harness its structured reasoning to solve queries where "just Google it" falls short.