Editorial illustration for Nvidia Says Nemotron 3.5 Super VL Outperforms Gemini Ultra in Agent Tests
Nvidia's Nemotron 3.5 Beats Google Gemini Ultra
Nvidia Says Nemotron 3.5 Super VL Outperforms Gemini Ultra in Agent Tests
Nvidia put a number on it Tuesday: Nemotron 3.5 Super VL, its new open multimodal model, beats Google's Gemini Ultra on agent benchmarks, according to the company's own technical blog post. The model carries 120 billion parameters but uses a Mixture-of-Experts setup that activates only 12 billion per token, a design meant to keep costs down while handling text, images, documents, and video in one pass. Nvidia built it for agents that run around the clock in enterprise settings rather than for one-off chat sessions.
The release isn't solo work. Nvidia developed Nemotron 3.5 Super VL with the Nemotron Coalition, a group that includes Mistral AI, Thinking Machines Lab, Perplexity, and Black Forest Labs. Alongside the flagship model, Nvidia is shipping two smaller Bolt variants aimed at local hardware: one for machines with 24 GB of memory or more, and a 4 GB version built for voice assistants that follows later this month. Nvidia is also publishing datasets and training recipes, not just weights, though everything is tuned to run best on its own chips.
Nvidia distilled Super VL from Nemotron 3 Ultra, its largest reasoning model. On tasks where an agent follows instructions, calls tools, works through multi-step business processes, or analyzes images, Super VL is supposed to match Ultra's accuracy with only about one-fifth the parameters.
Why this matters
Nvidia putting weights, datasets, and training recipes out in the open changes the calculus for anyone building agents right now. A 120-billion-parameter model that beats Gemini Ultra on Hermes and LangChain's agent benchmarks, and still clears Nemotron 3 Super on SWE-bench and Terminal-Bench, gives developers a credible open alternative to the closed APIs they've been locked into for coding and tool-use workloads. The Bolt variants matter just as much as the flagship: a model that fits in 24 GB of memory, with a 4 GB version aimed at voice assistants, means teams can run agents on local hardware instead of routing every call through a cloud provider.
That's a real cost and latency argument, not a marketing one. We'd still want independent benchmark runs before taking Nvidia's own comparisons at face value, since the company has every incentive to frame Nemotron favorably against Ultra. But if the throughput claims hold, five times the performance per GPU versus comparable open models, this is the kind of release that pulls enterprise agent development away from proprietary stacks and back toward infrastructure Nvidia controls end to end.
Related Reading
Browse all Google news →
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
From Archives openKylin Outlines Vision for Reimagining Linux in AI Era at Open Source Summit
From Archives NVIDIA and Sweden AI Factory Expand Compute Access for European SMEs
From Archives Google Launches 'Universal' Shopping Cart for Search and Gemini
Business & Startups
From Archives Google: self-improving AI far off as fully delegated code changes rose to 11.9%
Research & Benchmarks
Common Questions Answered
How does Nemotron 3.5 Super VL achieve better performance than Gemini Ultra while using fewer parameters?
Nemotron 3.5 Super VL uses a Mixture-of-Experts architecture that activates only 12 billion parameters per token, despite having 120 billion total parameters. This efficient design was distilled from Nemotron 3 Ultra and allows it to match Ultra's accuracy on agent benchmarks while using approximately one-fifth of the parameters, significantly reducing computational costs.
What types of tasks is Nemotron 3.5 Super VL optimized for in enterprise environments?
Nemotron 3.5 Super VL is specifically designed for agents that run continuously in enterprise settings and excel at instruction-following, tool-calling, multi-step business process execution, and image analysis. The model handles text, images, documents, and video in a single pass, making it well-suited for complex agent workflows.
Why does Nvidia releasing Nemotron 3.5 Super VL as an open model matter for developers?
By releasing the model weights, datasets, and training recipes openly, Nvidia provides developers with a credible open alternative to closed APIs for coding and tool-use workloads. This breaks the dependency lock-in that developers have experienced and gives them more flexibility and control over their agent implementations while maintaining competitive performance on benchmarks like Hermes and LangChain.
What benchmarks demonstrate Nemotron 3.5 Super VL's superiority over competing models?
Nemotron 3.5 Super VL outperforms Google's Gemini Ultra on agent benchmarks including Hermes and LangChain's agent benchmarks. Additionally, the model clears Nemotron 3 Super on SWE-bench and Terminal-Bench, demonstrating strong performance across multiple evaluation frameworks for agent capabilities.
Further Reading
- Papers with Code - Latest NLP Research - Papers with Code
- Hugging Face Daily Papers - Hugging Face
- ArXiv CS.CL (Computation and Language) - ArXiv
