DEV.co
Open-Source LLM · Alibaba-NLP

Tongyi-DeepResearch-30B-A3B

Tongyi-DeepResearch-30B is a 30-billion-parameter mixture-of-experts model from Alibaba that activates only 3B parameters per token. It is purpose-built for agentic tasks requiring long-horizon reasoning and information-seeking (e.g., web research, multi-step problem solving). The model is open-source under Apache 2.0, ungated, and compatible with ReAct and a heavier 'IterResearch' inference mode for test-time scaling.

Source: HuggingFace — huggingface.co/Alibaba-NLP/Tongyi-DeepResearch-30B-A3B
30.5B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
65.8k
Downloads (30d)

Key facts

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

FieldValue
DeveloperAlibaba-NLP
Parameters30.5B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads65.8k
Likes812
Last updated2025-10-10
SourceAlibaba-NLP/Tongyi-DeepResearch-30B-A3B

What Tongyi-DeepResearch-30B-A3B is

A Qwen3-based MoE architecture with 30.5B total parameters and selective activation (3B active per token). Trained via: (1) synthetic data generation pipeline for agentic pretraining, (2) large-scale continual pretraining on agentic interaction data, (3) end-to-end reinforcement learning using Group Relative Policy Optimization with token-level gradients. Inference supports ReAct (core ability evaluation) and IterResearch 'Heavy' mode (test-time scaling). Context length unknown. Compatible with Hugging Face transformers, safetensors, and Azure deployment.

Quickstart

Run Tongyi-DeepResearch-30B-A3B locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="Alibaba-NLP/Tongyi-DeepResearch-30B-A3B")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Automated research and information gathering

Designed for deep information-seeking tasks. Strong performance on agentic search benchmarks (BrowserComp, WebWalkerQA, GAIA, xbench-DeepSearch, FRAMES, Humanity's Last Exam). Suitable for building autonomous research agents and knowledge discovery systems.

Long-context question answering with reasoning

Excels at multi-step reasoning over extended reasoning chains. IterResearch 'Heavy' mode enables test-time compute scaling for higher-quality outputs on complex queries. Ideal for technical Q&A, document analysis, and scenario-based problem solving.

Self-hosted agentic AI applications

Ungated, Apache 2.0 licensed, and deployable on Azure and other cloud platforms. 3B active parameters per token reduce inference cost vs. dense 30B models. Suitable for cost-conscious organizations needing on-premises or private-cloud agentic agents.

Running & fine-tuning it

ESTIMATE: 60–120 GB GPU VRAM (unquantized, float16/bfloat16). MoE architecture with 3B active parameters may allow partial loading optimizations. Quantization (int8, int4) can reduce footprint to ~15–30 GB. Multi-GPU setups recommended for inference throughput. CPU inference not practical at this scale.

Model card does not explicitly discuss LoRA, QLoRA, or fine-tuning feasibility. Likely fine-tunable given Qwen3 base and Hugging Face transformers compatibility, but specific guidance missing. Recommend consulting GitHub repository (https://github.com/Alibaba-NLP/DeepResearch) for training scripts and adapter support. Reinforcement learning framework mentioned (GRPO) suggests RL fine-tuning is possible but may require specialized setup.

When to avoid it — and what to weigh

  • Need guaranteed real-time latency — IterResearch 'Heavy' mode uses test-time scaling (longer inference time for better quality). ReAct mode is faster but may trade depth. If sub-100ms latency is critical, validate performance empirically first.
  • Limited deployment infrastructure — Even with 3B active parameters, total model size (~30B) requires substantial GPU VRAM (estimate 60–120 GB unquantized depending on precision). Requires modern hardware or quantization. Limited suitable for edge or CPU-only deployments.
  • Out-of-domain tasks unrelated to search/research — Model is specifically optimized for agentic search and long-horizon reasoning. General-purpose tasks (classification, summarization, translation) may underperform compared to broader LLMs. Benchmark performance reported only on agentic tasks.
  • Production use without validation of benchmarks — Benchmarks (GAIA, BrowserComp, etc.) are stated but not detailed in the card. Actual performance on your specific task unknown. Requires thorough evaluation before production deployment.

License & commercial use

Apache License 2.0 (OSI-approved, permissive). Allows commercial use, modification, and distribution with proper attribution and liability disclaimer. No proprietary restrictions stated.

Apache 2.0 is permissive and explicitly allows commercial use. Model is ungated and freely downloadable. No evidence of restrictions on commercial deployment. However: (1) verify Alibaba's policy on derivative works; (2) ensure compliance with any downstream dependencies (e.g., training data sources); (3) test thoroughly for your use case before committing to production, as benchmark performance is agentic-specific and may not transfer.

DEV.co evaluation signals

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

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

No explicit security audit, data provenance statement, or adversarial robustness evaluation provided. Model trained on synthetic + real agentic interaction data; source and filtering criteria unknown. Agentic inference (web browsing, action execution) introduces risk of prompt injection and misuse; enforce guardrails at application level. Quantization may degrade robustness slightly. No known vulnerabilities stated. Requires risk assessment before deploying in high-stakes or public-facing systems.

Alternatives to consider

GPT-4o / Claude 3.5 Sonnet (proprietary)

Broader capabilities, stronger general reasoning, no deployment burden. Higher cost and closed-source. Better for organizations prioritizing capability over cost and control.

Qwen2.5-72B or Llama 3.1-70B (dense open-source)

Larger dense models with broader training. May outperform on non-agentic tasks and general reasoning. Higher VRAM footprint; no MoE efficiency. Consider if your workload is diverse.

Agent frameworks + smaller model (e.g., LangChain + Mistral 7B)

Lower hardware cost and latency. May require more orchestration and tool prompting to match Tongyi's agentic performance. Better for resource-constrained deployments.

Software development agency

Ship Tongyi-DeepResearch-30B-A3B with senior software developers

Ready to build autonomous research agents or deep-reasoning applications? Tongyi-DeepResearch is open-source, commercially permissive, and optimized for long-horizon information-seeking tasks. Start with the GitHub repository (Alibaba-NLP/DeepResearch) for inference scripts, or contact our team to architect a scalable deployment on your infrastructure.

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.

Tongyi-DeepResearch-30B-A3B FAQ

Can I use this model commercially?
Yes. Apache 2.0 is permissive and explicitly permits commercial use, modification, and distribution with attribution. No gatekeeping or proprietary license. However, validate that training data sources and any downstream integrations comply with your legal requirements.
What GPU do I need to run this?
Estimate 60–120 GB VRAM (unquantized, float16/bfloat16). With quantization (int4), ~15–30 GB. Multi-GPU recommended for throughput. A100/H100 or RTX 6000 series. Context length unknown; longer sequences will increase memory. Test on your target hardware before production.
How do I fine-tune it?
Model card does not specify. Likely compatible with transformers LoRA/QLoRA given Qwen3 base, but no explicit guidance. See GitHub repository (Alibaba-NLP/DeepResearch) for training scripts and adapters. GRPO framework mentioned suggests RL fine-tuning possible but may require custom setup.
Is this suitable for my non-agentic use case (e.g., classification, summarization)?
Unknown. Model is optimized for agentic search and long-horizon reasoning. Benchmarks are agentic-specific. Performance on general NLP tasks not reported. Recommend evaluation on your specific task before adoption.

Custom software development services

From first prototype to production, DEV.co delivers software development services around tools like Tongyi-DeepResearch-30B-A3B. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across open-source llms and beyond.

Deploy Tongyi-DeepResearch for Your Agentic AI Stack

Ready to build autonomous research agents or deep-reasoning applications? Tongyi-DeepResearch is open-source, commercially permissive, and optimized for long-horizon information-seeking tasks. Start with the GitHub repository (Alibaba-NLP/DeepResearch) for inference scripts, or contact our team to architect a scalable deployment on your infrastructure.