Fine-tune an AI agent for network failures: NVIDIA offers a Qwen3-32B recipe

An open model can be trained to choose how to resolve a network failure using examples from your own incidents.

In NVIDIA's training recipe dated Feb 27, fine-tuned Qwen3-32B selects the correct solution for 19 of 21 synthetic incidents, up from 17.

The base model answered 81.0% correctly. After training, NVIDIA expects 90.5% on the same training set. The recipe prepares an AI agent for a network operations center: it analyzes an incident and selects a fix. Training examples are prepared using gpt-oss-120b, which generates reasoning for the solutions.

Adapt the recipe to your data. Clone the `NVIDIA-NeMo/Skills` repository, install the package with `pip install .`, go to `recipes/noc-reasoning-agent` and run `ns setup`. Scripts, configs and sample data are included. Replace `data/synthetic_incidents.csv` with your own incidents, keeping the same column schema.

The full data preparation, training and evaluation cycle requires 8 NVIDIA A100 GPUs or newer and takes several hours. Docker, NVIDIA Container Toolkit and Python 3.10+ are required.

A ready-made industry model. On Feb 28, NVIDIA also introduced the 30-billion-parameter Nemotron LTM, which AdaptKey fine-tuned on open telecom data, including industry standards and synthetic logs. AdaptKey-Nemotron-30b is available on Hugging Face. Downloading it requires accepting the terms and agreeing to share contact details.

AdaptKey used LoRA fine-tuning on 1.3 million industry examples. In the model card as of Oct 6, the company reports its own TeleFlow measurements:

| Model | Total TeleFlow score | | --- | ---: | | Base Nemotron-3-Nano-30B-A3B | 538 | | Fine-tuned AdaptKey-Nemotron-30b | 596 |

The total score on this task set improved by 10.8%. The model card includes examples for running it with Transformers and vLLM, as well as Docker Compose with an API on port 8090.

Operators also use open models for their own projects. Rajiv Koodli of SoftBank describes a Large Telecom Model based on Nemotron and other models for network operations, design and management. Chirag Sukhadia of Indosat Ooredoo Hutchison describes another use case: the Sahabat-AI family adapts models to Indonesia's language, culture and local applications.

Primary sources: [NVIDIA on operators' strategies, Oct 6, 2026](https://blogs.nvidia.com/blog/telecom-operators-open-models/), [AdaptKey-Nemotron-30b model card with measurements and deployment examples](https://huggingface.co/AdaptKey/AdaptKey-Nemotron-30b).

In the State of AI in Telecommunications report, which NVIDIA cites on Oct 6, 89% of respondents consider open models and software important to their company's AI strategy.