← Back to feed

NVIDIA’s AI-Q is not a product. It is a reference architecture — a blueprint for building enterprise research agents using fully open, self-hostable components. On the two benchmarks that matter most in the deep research space, DeepResearch Bench I and DeepResearch Bench II, it currently ranks first. That is the same benchmark suite where OpenAI’s closed Deep Research system has been the standard to beat.

The distinction matters. OpenAI Deep Research cannot be inspected, swapped, or deployed inside your own infrastructure. AI-Q can.

Architecture

AI-Q is built on NVIDIA’s NeMo Agent Toolkit and LangChain Deep Agents, orchestrated through LangGraph. The system operates in two modes:

Shallow research — a bounded, fast path for queries that need a few tool calls and a concise cited answer. Returns in seconds.

Deep research — a multi-phase planner that dispatches parallel specialist researcher agents, iterates over evidence gaps, and produces a structured report with a table of contents, inline citations, and cross-referenced sources. A full deep research run produces output comparable to a professional analyst brief.

The entire routing logic — which mode fires, which model handles which agent, what tools are in scope — is defined in YAML config files. There is no code change required to swap models, add data sources, or adjust research depth thresholds. Every component (orchestrator, planner, researcher, individual sub-agents) can run a different LLM.

Deployment

AI-Q deploys via Docker Compose or Helm charts. It supports NVIDIA Nemotron models via NVIDIA NIM for fully offline, air-gapped operation. A Dell-NVIDIA AI-Q 2.0 Reference Architecture is available for production deployments on Dell AI Factory hardware.

The latest version adds first-class support for authenticated MCP servers as data sources, which means research pipelines can pull from enterprise knowledge bases, internal wikis, and proprietary data stores without routing through public APIs.

Plugs Into Existing Agent Harnesses

AI-Q ships as a packaged agent skill under .agents/skills/aiq-research/ in the GitHub repository. A SKILL.md interface file and a scripts/aiq.py helper handle routing, job submission, polling, and result retrieval. Compatible out of the box with Codex, Claude Code, and OpenCode as a delegated deep research backend — the harness submits the query, AI-Q runs the pipeline asynchronously, and returns a structured cited report.

The Benchmark Position

DeepResearch Bench I and II test multi-source synthesis, citation accuracy, long-horizon reasoning, and report structure. OpenAI’s Deep Research had been the strongest performer on this benchmark family. AI-Q, built entirely from open components, now leads both.

The practical implication: organisations that cannot route sensitive queries through third-party APIs — financial services, defence contractors, healthcare systems — have a production-grade research agent that runs under their own governance. NVIDIA is not selling a service here; it is commoditising the pipeline.

The GitHub repository is public under the NVIDIA-AI-Blueprints organisation.