As enterprises shift from simple chatbots to multi-agent systems, underlying AI models are facing unprecedented performance and cost challenges. To address these pain points, NVIDIA announced the new Nemotron 3 Super model, an open-weighted model with 1200 billion parameters and a hybrid expert (MoE) architecture. By providing a massive context window of up to 100 million tokens and being deeply optimized for NVIDIA's next-generation Blackwell architecture computing platform, Nemotron 3 Super not only increases data throughput by 5 times but also accurately solves the problems of "context inflation" and "thinking tax" in complex agent workflows.
Two major constraints on proxy AI: Context inflation and the tax on thinking
Why are existing large language models (LLMs) struggling to handle complex proxy tasks? NVIDIA points out two major development bottlenecks currently facing enterprises:
• First of allContext inflationIn collaborative workflows involving multiple AI agents, the system must continuously exchange complete historical records, tool outputs, and intermediate reasoning processes. This results in the generation of more than 15 times the number of lexical units compared to typical conversational interactions. This massive amount of data not only drives up computational costs but also frequently causes AI to "forget" or deviate from its original goal when handling lengthy tasks.
• followed by"Thinking Tax"A competent autonomous agent must perform deep reasoning at every step of the task execution. However, if every tiny subtask requires calling a massive model with hundreds of billions of parameters, the application will run extremely slowly and be prohibitively expensive, making it impossible to deploy on a large scale in an enterprise environment.
Hybrid architecture unleashes its power: the ultimate performance of Mamba combined with Transformer
To address these issues, Nemotron 3 Super features a massive context window of 1 million tokens, allowing agents to retain the complete workflow state in memory. At the underlying architecture level, NVIDIA has gone all out, introducing three key innovations that result in a 5x increase in data throughput and a 2x increase in accuracy compared to its predecessor:
• Hybrid Architecture:Breaking away from the myth of a single architecture, Nemotron 3 Super cleverly combines two neural networks. The Mamba layer provides up to 4 times the memory and computational efficiency (especially suitable for processing very long texts), while the traditional Transformer layer drives complex high-order inference.
• Advanced hybrid expert models and potential hybrid expert models:Although the model has a total of 120 billion parameters, only 1200 billion active parameters are activated at a time during the inference phase, significantly reducing the computational burden. Even more groundbreaking is the "Latent Hybrid Expert Model" (Latent MoE) technology, which can predict the next word with "the computational cost of one expert, while activating four expert models" during inference, thereby squeezing out higher accuracy without increasing computing power.
• Multi-Token Prediction: Breaking the previous limitation of only being able to utter one word at a time, the model can simultaneously predict multiple future words, increasing the overall reasoning speed by 3 times.
Optimized for the Blackwell architecture, fully open source to support the ecosystem.
Beyond its innovative software architecture, Nemotron 3 Super is a powerful demonstration of NVIDIA's capabilities specifically designed for the Blackwell GPU platform. On the Blackwell architecture platform, this model can run in the extremely low-precision NVFP4 format, making its inference speed up to four times faster than the previous generation Hopper architecture platform (running in FP8), without sacrificing accuracy.
In terms of open source attitude, NVIDIA has been extremely generous this time. The Nemotron 3 Super not only releases open weights with a permissive licensing approach, but also completely discloses its training dataset of more than 10 trillion words, 15 reinforcement learning environments, and complete evaluation process research methods.
Currently, companies including Perplexity, Amdocs, Palantir, Dassault Systèmes, and Siemens have begun deploying the Nemotron 3 Super model to drive internal software development or vertical domain automation. Enterprise developers can access this NVIDIA NIM microservice starting today through build.nvidia.com, Hugging Face, or major public cloud platforms such as Google Cloud, Oracle, and Microsoft Azure.
Analysis of viewpoints
The launch of the Nemotron 3 Super once again proves that NVIDIA is not just a "hardware company that sells chips".
While OpenAI and Anthropic were still arguing over subscription fees for closed-source models, NVIDIA chose a completely different strategy: "Give you the best software and models for free, as long as you continue to buy my hardware."
The most formidable aspect of the Nemotron 3 Super lies in its "complete optimization for NVIDIA's own hardware." By addressing the memory consumption issues associated with long texts through a hybrid "Mamba + Transformer" architecture, and leveraging the precision of NVFP4 to capitalize on the computing power of Blackwell GPUs, NVIDIA is essentially setting a standard for the integrated hardware and software of future "Agentic AI." The complete release of its 10-megabyte training dataset is a bombshell for the open-source community and will significantly accelerate the transition of enterprise-level AI agents from the lab to real-world production lines.
However, the real game-changer might be the rumored AI agent application that will be announced at GTC 2026, focusing on enterprise-level AI agent applications.NVIDIA's version of the green lobster, "NemoClaw"This technology could potentially break down hardware silos, allowing enterprises to seamlessly integrate even if their underlying AI chips don't use NVIDIA's proprietary chips. It appears this technology is already being rolled out to enterprise software giants such as Salesforce, Cisco, Google, Adobe, and CrowdStrike, with specific details expected to be revealed at GTC 2026.
