NVIDIA's 49B Nemotron Just Topped the DeepResearch Bench—and It Runs on a Single GPU

The open-source agent stack just got a new leader. NVIDIA’s AI-Q blueprint, a portable deep research agent, has climbed to the top spot on the Hugging Face “LLM with Search” leaderboard, scoring a 40.52 on the DeepResearch Bench. That benchmark isn’t a tidy multiple-choice quiz—it’s 100+ real-world research tasks requiring multi-hop reasoning and report-length synthesis across science, finance, and history.

What makes this notable isn’t just the score. It’s the hardware requirements. The stack pairs Meta’s Llama 3.3-70B Instruct with a smaller, optimized variant: Llama-3.3-Nemotron-Super-49B-v1.5. That 49B model was built using neural architecture search and knowledge distillation, and it can run on a single H100 GPU. For developers who’ve been struggling with the cost and complexity of deploying large reasoning models, that changes the math considerably.

The Nemotron model has a feature I haven’t seen widely discussed: a reasoning ON/OFF toggle. You can run it as a standard chat model or flip it into deep, chain-of-thought reasoning mode via a system prompt. That kind of granular control is genuinely useful when you’re trying to balance cost, latency, and depth of analysis in a production pipeline. It’s not one-size-fits-all thinking.

Transparency is the other thread worth pulling on. NVIDIA is making the post-training data, evaluation methods, and model lineage openly traceable from Meta’s original weights. In an ecosystem where “open” often means a model card and a shrug, that’s a meaningful distinction. The question now is whether the community can replicate these results outside of NVIDIA’s carefully tuned reference architecture—or if this is just another benchmark victory that dissolves under real-world messiness.