DEV.co
AI Frameworks · Alibaba-NLP

DeepResearch

Tongyi DeepResearch is an open-source agentic LLM with 30.5B parameters (3.3B active) designed for autonomous web research and information-seeking tasks. It combines continual pre-training on agentic data, reinforcement learning, and inference compatibility with ReAct and IterResearch paradigms to achieve state-of-the-art performance on search benchmarks.

Source: GitHub — github.com/Alibaba-NLP/DeepResearch
19.6k
GitHub stars
1.5k
Forks
Python
Primary language
Apache-2.0
License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
RepositoryAlibaba-NLP/DeepResearch
OwnerAlibaba-NLP
Primary languagePython
LicenseApache-2.0 — OSI-approved
Stars19.6k
Forks1.5k
Open issues91
Latest releaseUnknown
Last updated2026-02-27
Sourcehttps://github.com/Alibaba-NLP/DeepResearch

What DeepResearch is

Built on Alibaba's Tongyi Lab, the model uses a synthetic data generation pipeline for agentic pre-training and SFT, employs Group Relative Policy Optimization (GRPO) for on-policy RL with token-level gradients and leave-one-out advantage estimation, and supports both lightweight ReAct and heavy test-time scaling inference modes. Context length: 128K tokens.

Quickstart

Get the DeepResearch source

Clone the repository and explore it locally.

terminalbash
git clone https://github.com/Alibaba-NLP/DeepResearch.gitcd DeepResearch# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Autonomous Multi-step Web Research

Perform deep, long-horizon information-seeking tasks that require iterative web search, document retrieval, and synthesis—particularly for comprehensive topic analysis, competitive intelligence, and market research without manual intervention.

Production Agent Deployment

Deploy as a backbone for agentic systems requiring deterministic reasoning and tool use; Apache 2.0 license enables commercial productization. Available via HuggingFace, ModelScope, OpenRouter, and Alibaba's Bailian service for managed inference.

Benchmark Validation & Research

Evaluate against established agentic search benchmarks (Humanity's Last Exam, BrowseComp, WebWalkerQA, SimpleQA) with published results; suitable for academic teams validating new research in agent design or comparing with open alternatives.

Implementation considerations

  • Requires Python 3.10.0 specifically; dependency conflicts reported with other versions per README; isolate environment via conda or virtualenv.
  • Mandatory API key provisioning: Serper (web search), Jina (page reading), OpenAI-compatible endpoint (summarization), Dashscope (file parsing), SandboxFusion (code execution). Missing keys will degrade capabilities.
  • Inference split: lightweight ReAct mode for reproducibility vs. heavy IterResearch mode for maximum performance; choose based on latency tolerance and output quality requirements.
  • Model size (30B-A3B, 128K context) demands GPU with minimum 24GB VRAM for inference; exact quantization/batching strategies and memory requirements not detailed in README.
  • Evaluation data format: JSONL or JSON with `question` and `answer` fields; file references require explicit filename prepends; ground-truth answers used for benchmark comparison, not generation guidance.

When to avoid it — and what to weigh

  • Lightweight Edge Deployment — 30B parameters require significant GPU VRAM even with MoA quantization; not suitable for constrained environments like mobile, embedded systems, or inference on consumer hardware without specialized optimization.
  • Proprietary Data Protection — Model requires external tool calls (web search, page reading, code sandbox) for full capability; if you cannot expose queries to third-party services (Serper, Jina, Dashscope) due to confidentiality, local-only solutions are required.
  • Real-time Low-Latency Responses — README notes demo response times 'may vary or fail intermittently due to model latency and tool QPS limits'; agentic inference inherently adds multi-step latency unsuitable for sub-100ms SLA applications.
  • Non-English or Multilingual-Primary Use — Limited multilingual evidence; BrowseComp-ZH benchmark suggests Chinese support, but primary training data composition and tokenizer language coverage are not clearly documented.

License & commercial use

Apache License 2.0 (Apache-2.0). Permissive OSI-approved license permitting commercial use, modification, and distribution with liability disclaimer and trademark restrictions. No additional commercial restrictions observed in repository.

Apache 2.0 explicitly permits commercial use. Model weights available via HuggingFace and ModelScope for self-hosting; Alibaba also provides Bailian managed service. No proprietary model calls or usage restrictions detected, but production deployments should be tested independently for compliance and performance SLAs.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceHigh
Security considerations

Model executes external tool calls (web search, code sandbox); queries exposed to third-party services (Serper, Jina, Dashscope, SandboxFusion). Code sandbox from ByteDance SandboxFusion; security posture of that component not assessed here. Local deployment mitigates third-party exposure but requires secure API key management via `.env` files (gitignored). No security audit or vulnerability disclosure process documented.

Alternatives to consider

OpenAI o1 / GPT-4 with Agents

Closed-source, proprietary, battle-tested in production; stronger reasoning but no local control, higher per-query cost, and vendor lock-in.

Llama 3.1 (70B/8B) + ReAct framework

Fully open-source, larger parameter scale available, but requires manual agentic framework integration; no built-in agent pre-training or RL tuning like DeepResearch.

Claude API with Tools

Closed-source, proprietary; strong on multi-step reasoning and tool use, but less transparency, higher cost, and no local deployment option.

Software development agency

Build on DeepResearch with DEV.co software developers

Start with HuggingFace or Alibaba Bailian for managed inference, or self-host using our local deployment guide. Evaluate on your benchmarks and contact our team for production SLA discussion.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

DeepResearch FAQ

Can I use DeepResearch for real-time applications?
README explicitly warns that demo response times 'may vary or fail intermittently'; agentic multi-step inference adds inherent latency. Suitable for research and background tasks, not real-time or low-latency SLAs.
What GPU hardware is required?
Not explicitly specified in provided data. 30B parameters typically require 24GB+ VRAM for inference; exact quantization strategies and memory breakdowns are not documented. Requires independent benchmarking for your infrastructure.
Is there a smaller or quantized version?
Only 30B-A3B variant is listed. No 4-bit, 8-bit, or smaller distilled models documented in README. Community quantization (GGUF, etc.) would depend on HuggingFace ecosystem contributions.
Can I fine-tune the model on proprietary data?
Apache 2.0 license permits modification. Practical fine-tuning methodology (LoRA, QLoRA, full SFT) not detailed in README; consult accompanying paper or community forums for training scripts.

Custom software development services

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If DeepResearch is part of your ai frameworks roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy Agentic Research?

Start with HuggingFace or Alibaba Bailian for managed inference, or self-host using our local deployment guide. Evaluate on your benchmarks and contact our team for production SLA discussion.