One number defines the problem: 1,500 percent. That is how much more tokens advanced multi-agent AI workflows generate compared to standard formats, according to the announcement. Every interaction forces the system to resend full conversation histories, intermediate reasoning chains, and tool outputs. Multiply that across extended enterprise tasks, and costs compound fast — while agents themselves begin to drift from their original objectives.
Two constraints define the economics of building with multi-agent AI at scale. The first is what the announcement calls the “thinking tax” — the computational cost of requiring large architectures to reason through every subtask independently. The second is context explosion, the token volume problem that both inflates infrastructure spending and introduces goal drift, where autonomous agents diverge from their assigned objectives over long workflows.
NVIDIA is positioning its new open-weights model, Nemotron 3 Super, as a direct answer to both. The architecture carries 120 billion parameters, but only 12 billion remain active during inference. That distinction matters enormously for enterprise cost calculations.
How the architecture manages the efficiency gap
The model uses a hybrid mixture-of-experts design built on three interlocking mechanisms. Mamba layers deliver four times the memory and compute efficiency. Standard transformer layers handle complex reasoning. A latent technique activates four expert specialists for the cost of one during token generation. Taken together, the system delivers up to five times higher throughput and twice the accuracy of its predecessor, the earlier Nemotron Super model.
Inference speed gets a separate boost through speculative decoding — the system anticipates multiple future words simultaneously, accelerating output threefold. Running on NVIDIA‘s Blackwell platform with NVFP4 precision, the architecture cuts memory requirements and processes inference up to four times faster than FP8 configurations on Hopper systems, without accuracy loss.
The context window reaches one million tokens, allowing agents to hold an entire workflow state in memory at once. A software development agent can load a full codebase simultaneously, enabling end-to-end code generation and debugging without segmenting documents. In financial analysis, thousands of pages of reports load into a single context, removing the need to re-reason across fragmented conversations.
Where it is already being deployed
The early adoption list is specific. Amdocs, Palantir, Cadence, Dassault Systèmes, and Siemens are customising the model across telecom, cybersecurity, semiconductor design, and manufacturing. Software development platforms CodeRabbit, Factory, and Greptile are integrating it alongside their own proprietary models, targeting higher accuracy at lower operational cost.
Life sciences firms Edison Scientific and Lila Sciences will use it to power agents handling deep literature search, data science tasks, and molecular analysis. The model also pushed the AI-Q agent to the top position on the DeepResearch Bench and DeepResearch Bench II leaderboards, which measure multistep research performance across large document sets.
NVIDIA released the model with open weights under a permissive license. It is packaged as an NVIDIA NIM microservice, deployable across workstations, data centres, and cloud environments. The model was trained on synthetic data and claimed the top spot on Artificial Analysis for efficiency and openness among models of its parameter class.
Photo by Brett Sayles on Pexels
This article is a curated summary based on third-party sources. Source: Read the original article
