Skip to main content
Back

Alibaba's Tongyi-DeepResearch Beats OpenAI o3 at Web Research With 3B Active Parameters

Alibaba's Tongyi Lab released an open-source 30B MoE agent that activates just 3B parameters and tops deep-research benchmarks against o3.

  • Alibaba open-sourced Tongyi-DeepResearch-30B-A3B, a 30B MoE agent activating only 3B parameters per token.
  • Beats OpenAI o3 on HLE (32.9 vs 24.9) and xbench (75.0 vs 67.0) benchmarks.
  • Scores 90.6 on FRAMES, 70.9 on GAIA, 43.4 on BrowseComp with 128K context.
  • Two inference modes: ReAct for standard use, IterResearch Heavy for max accuracy.
  • Trained via automated synthetic data, continual pretraining, and on-policy GRPO reinforcement learning.
  • Apache-2.0, runs on vLLM, SGLang, Ollama; code on GitHub.

Alibaba's Tongyi Lab has released Tongyi-DeepResearch-30B-A3B, an open-weights agent model built for multi-step web research. The architecture is a sparse Mixture-of-Experts with 30.5 billion total parameters but only 3.3 billion activated per token, keeping inference costs manageable. The release is Apache-2.0 licensed, hosted on Hugging Face, and ships with an inference stack on GitHub.

The goal is to match or beat the closed deep-research systems from OpenAI and Google at a fraction of the compute cost.

Scoreboard: beating models several times its size

On benchmarks designed for agentic search, Tongyi-DeepResearch-30B-A3B scores 32.9 on Humanity's Last Exam, 43.4 on BrowseComp, 70.9 on GAIA, 75.0 on xbench-DeepSearch, and 90.6 on FRAMES. It outperforms OpenAI o3 on HLE (32.9 vs. 24.9) and xbench (75.0 vs. 67.0) despite o3's larger scale, and claims 35.2 on BrowseComp-ZH, a 10-point lead over Gemini. The model also exposes a 131,072-token context window, relevant when an agent needs to accumulate dozens of scraped pages inside a single trajectory.

Built to browse, not chat

Tongyi-DeepResearch is tuned to plan searches, call tools, read pages, and iterate across long research sessions. Teams are already deploying it for legal research (parsing statutes and case law with citations), multi-day travel planning against real-time data, and powering Alibaba's Gaode Mate navigation app to query flights, hotels, and events in parallel. Any workflow requiring verifiable, cited answers drawn from many sources fits the use case better than a one-shot generation model would.