NVIDIA AI-Q Tops DeepResearch Bench I and II
Key Facts
- What: NVIDIA AI-Q deep research agent achieved first place on DeepResearch Bench I with a score of 55.95 and DeepResearch Bench II with 54.50.
- Architecture: Multi-agent system consisting of planner, researcher, and orchestrator built on NVIDIA NeMo Agent Toolkit and LangChain DeepAgents, powered by fine-tuned NVIDIA Nemotron 3 Super models.
- Approach: Open, modular blueprint that is fully configurable and portable for enterprise use cases, combining web search (Tavily), academic search (Serper), and citation-backed reporting.
- Key Techniques: Custom fine-tuning of Nemotron-3-Super-120B-A12B on roughly 67k SFT trajectories, custom middleware for long-horizon reliability, and optional ensemble + report refiner.
- Significance: Demonstrates that open, developer-accessible models and tooling can achieve state-of-the-art performance in agentic deep research.
NVIDIA AI-Q, developed in collaboration with Hugging Face, has taken the top spot on both primary benchmarks for deep research agents, according to a new technical blog post. The open blueprint for building AI agents reached scores of 55.95 on DeepResearch Bench I and 54.50 on DeepResearch Bench II, outperforming other systems in comprehensive report quality and granular analytical accuracy.
The achievement highlights the potential of fully open and modular architectures that enterprises can inspect, customize and deploy for their specific needs. Unlike closed systems, AI-Q provides a transparent stack built on NVIDIA technologies and open frameworks, offering a reproducible path to advanced agentic research capabilities.
Benchmark Performance
DeepResearch Bench I and II evaluate research agents through complementary lenses. Bench I assesses overall report quality against reference reports across dimensions including comprehensiveness, depth of insight, instruction-following and readability. Bench II employs over 70 fine-grained binary rubrics per task to measure information recall, analytical synthesis and clear presentation of findings.
By leading on both benchmarks, the NVIDIA AI-Q deep researcher demonstrates strength in producing polished, well-cited reports while maintaining strong underlying retrieval and reasoning, according to the Hugging Face blog post published March 12, 2026.
Technical Architecture
The AI-Q deep researcher follows a multi-agent architecture with three core components: an orchestrator that manages the overall research loop, a planner that creates an evidence-grounded research plan, and a researcher that deploys parallel specialist agents across multiple analytical perspectives. Each agent can be powered by different large language models.
An optional ensemble approach runs multiple agent pipelines in parallel and merges their outputs to maximize report quality and information coverage. The system is built using the NVIDIA NeMo Agent Toolkit for workflow composition and LangChain DeepAgents for the multi-phase planner-researcher-orchestrator flow.
The stack relies on fine-tuned NVIDIA Nemotron 3 models, which can be served through NVIDIA NIM or Build. Core capabilities include multi-step research processes, integration with Tavily for web search and Serper for academic papers, and generation of reports with proper citations.
Training and Optimization
A significant contributor to the results was the custom fine-tuning of the NVIDIA Nemotron-3-Super-120B-A12B model. The training used approximately 67,000 supervised fine-tuning (SFT) trajectories derived from seed datasets containing research questions. These were filtered using a principle-based judge to ensure quality.
The team also developed custom middleware extensions to the NeMo Agent Toolkit and LangChain components to improve long-horizon reliability and robustness. An optional ensemble layer combined with a post-hoc report refiner further enhanced output quality.
The full AI-Q blueprint extends beyond deep research to include intent routing, query clarification and shallow research capabilities, making it a comprehensive framework for enterprise AI agent development.
Impact on Open AI Research
This result represents a meaningful step for open, portable deep research systems. The fact that a single configurable, open-source-oriented stack leads both major benchmarks suggests that developer-accessible models and tooling can compete at the highest levels of agentic AI performance.
For developers and enterprises, AI-Q offers a practical reference architecture that can be owned and customized rather than treated as a black box. The open nature of the components allows for inspection, modification and optimization for specific industry or organizational requirements.
The achievement comes amid growing industry interest in sophisticated research agents capable of synthesizing information from large document collections while maintaining coherent reasoning across multiple steps.
What's Next
The Hugging Face blog post positions AI-Q as an open blueprint that the community can build upon. NVIDIA and its collaborators have made the architecture and methodologies available for reproduction and extension.
Future work may focus on further refinements to the fine-tuning process, additional middleware improvements, and expanding the range of use cases supported by the broader AI-Q framework. The modular design allows different organizations to swap in alternative models or tools as needed.
As benchmarks for deep research agents continue to evolve, maintaining leadership on both quality and analytical rigor will likely require ongoing innovation in multi-agent coordination and synthesis capabilities.
Sources
- How NVIDIA AI-Q Reached #1 on DeepResearch Bench I and II
- NVIDIA AI-Q Tops DeepResearch Bench Research Leaderboard
- NVIDIA AI-Q Achieves Top Score for Open, Portable AI Researcher
- Measuring Open-Source Llama Nemotron Models on DeepResearch Bench
All technical specifications, pricing, and benchmark data in this article are sourced directly from official announcements. Competitor comparisons use publicly available data at time of publication. We update our coverage as new information becomes available.