NVIDIA Nemotron 3 Super 120B: Testing on Real Analytics Tasks on a Single GPU

NVIDIA released Nemotron 3 Super 120B — a model with 120 billion parameters that, according to the company, fits entirely on a single graphics card while maintaining a context window of 256 thousand tokens. The Luxms BI team — a Russian business analytics platform — decided to test these claims in practice and spent a week testing on real corporate data. These are not synthetic benchmarks or demonstration examples: the model was integrated into a working BI tool and tested on live tasks.

A bit about the model itself. Nemotron 3 Super 120B is the flagship of NVIDIA's new model stack that launched in 2025. The architecture is optimized for efficient inference on single GPUs without needing to assemble multi-card clusters.

Three parameters distinguish it among competitors in the open-weight class. The first — 120 billion parameters, comparable to the best open models. The second — a context window of 256 thousand tokens, one of the highest metrics in its class.

The third — native support for agentic behavior: the model can independently plan multi-step tasks, call external tools, and correct its actions based on intermediate results without constant operator involvement. The key question the Luxms BI team posed to itself: can real analytical tasks be addressed on a single GPU today where previously a cluster was required? This matters for two reasons.

First, economics: a GPU cluster in a corporate environment means significant capital and operational costs, especially for mid-sized companies. Second, security: many organizations cannot or fundamentally do not want to send sensitive data to cloud services. One powerful graphics card with a sufficient model represents different economics and a fundamentally different level of data control.

Testing covered three categories of tasks. The first — SQL query generation from natural language descriptions: an analyst describes what they want to see in a report, the model writes a query to a relational database. The second — dashboard interpretation: explaining metric dynamics, finding anomalies, identifying trends in numerical data.

The third and most informative — agentic analysis of multi-step business scenarios, where the model needed to sequentially access multiple data sources, compare results, and formulate an analytical conclusion. The 256K token context turned out to be not just a marketing figure but a practically significant parameter. In real analytics, it's often necessary to keep in memory simultaneously a large data schema, a chain of intermediate results, and broad business context.

Models with smaller context windows lose the thread of reasoning on long chains. Nemotron 3 Super handled this noticeably more robustly. Agentic mode confirmed potential but revealed a practical limitation: it requires careful tuning of system prompts and correctly organized tool environment.

Weaknesses also emerged. In tasks requiring deep domain expertise — financial analysis with industry specificity or multi-level logical reasoning with complex dependencies between factors — the model still lags behind the best proprietary solutions. This is an important reference point for those considering Nemotron as a complete replacement for cloud APIs.

Precise benchmarks for each category with numbers and specific examples — in the full version of the material. The overall market conclusion: the emergence of a competitive 120-billion-parameter model running on a single GPU is a significant shift in the accessibility of powerful language models for corporate analytics. Companies now have a real opportunity to deploy a productive model within their own infrastructure, without cloud services and without an expensive cluster.

NVIDIA systematically occupies positions not only in chip manufacturing but also in the model stack — and Nemotron 3 Super becomes a weighty argument for this strategy in the corporate b2b market.