NVIDIA releases a 30-billion-parameter Nemotron 3 Large Telco Model with a full fine-tuning recipe

In a blog post, NVIDIA argues that telecom operators are increasingly building their AI strategies on open models. It cites its latest State of AI in Telecommunications report, in which 89% of respondents said open source models and software are important to their company's AI strategy, and stresses that the value goes beyond cost alone.

NVIDIA lists five reasons: access to frontier-level intelligence at lower cost; telco-specific customisation using network, customer and industry data; greater visibility into and control over model artefacts for governance and compliance; flexible deployment across public cloud, private infrastructure and the edge; and the ability to host and fine-tune open models to serve enterprise and government customers locally.

Among those open foundations, NVIDIA highlights its Nemotron family, which it says offers frontier-level reasoning for agentic workflows plus speech capabilities, with open weights, training data and recipes. SoftBank Corp. is cited as an example.

Rajeev Koodli, principal fellow at SoftBank Corp. and senior vice president of SB Telecom America, said open models let SoftBank build on the rapid progress of global foundation models while applying its own network knowledge and operational expertise, and that the company is using open foundations including NVIDIA Nemotron models while developing its SoftBank Large Telecom Model for network operations, design and overall management.

More concretely, NVIDIA announced a 30-billion-parameter Nemotron 3 Large Telco Model, fine-tuned by AdaptKey on open source telecom datasets so it can understand industry terminology and reason through workflows such as network configuration and customer incident triage, giving operators an open baseline.

To let operators adapt that model and others with their own operational data, NVIDIA also released a full recipe describing an end-to-end fine-tuning pipeline built on its open NeMo libraries, covering how to tailor open models to an operator's own networks, customers and procedures.

Handing an industry as regulated and operationally sensitive as telecom to open models is the core of NVIDIA's pitch. Operators want frontier capability without surrendering network data and operational logic to a closed API, and open weights plus a tuning recipe speak directly to that trade-off.

What to watch next is whether the accuracy gains claimed for Nemotron 3 LTM on real network operations can be reproduced by third parties, and how many operators will fine-tune on their own infrastructure rather than keep buying managed services.