#1414th of 14 in Open models
deepresearch
Open 30B-A3B web research agent model with ReAct inference scripts
Live demo ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 20k
- License
- Apache-2.0
- Last commit
- Feb 2026
- Language
- Python
Missed the latest GitHub check; these facts may be older.
Overview
Tongyi DeepResearch is a 30.5B-parameter mixture-of-experts model (3.3B active per token, 128K context) trained for long-horizon web information seeking. The repo provides inference scripts that run it in ReAct or an IterResearch-based Heavy mode, plus benchmark evaluation scripts. Weights are on HuggingFace and ModelScope, and the model is also on OpenRouter.
Who it is for: Researchers and engineers building or evaluating web search agents
Strengths
- Open weights on HuggingFace and ModelScope, 128K context, 3.3B parameters active per token
- Two inference modes: ReAct and a test-time-scaling Heavy mode
- Includes evaluation scripts for agentic benchmarks such as BrowseComp and HLE
- Can be called through OpenRouter without a local GPU
Weaknesses
- Needs several external API keys: Serper, Jina, an OpenAI-compatible endpoint, DashScope
- Python interpreter tool requires a separate SandboxFusion endpoint
- Research scripts, not a packaged app; no Docker or web UI
- OpenRouter use requires editing react_agent.py; README gives no hardware requirements
What it needs
- GPU optional
- Needs Serper, Jina, OpenAI-compatible API, DashScope, SandboxFusion
- Models: Tongyi-DeepResearch-30B-A3B, OpenRouter
Also in Open models
See all 14| Rank | Project | Score |
|---|---|---|
| 1 | llmfitRust tool that checks which LLMs fit your RAM, CPU and GPU | 84 out of 100 |
| 2 | layaSingle-pass multilingual classifier for typed choice, score and yes/no decisions | 79 out of 100 |
| 3 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 72 out of 100 |
| #4 | unlimited-ocrOCR model that parses long documents and multi-page PDFs in one pass | 62 out of 100 |
| #5 | insightfaceFace detection, recognition and alignment toolbox with ArcFace and RetinaFace models | 57 out of 100 |