What specific performance improvements were demonstrated by the fine-tuned telco model?

Evaluation of the fine-tuned Qwen3-32B model demonstrated significant accuracy gains, with performance for incident summary prediction and root-cause resolution increasing from a baseline of approximately 20% to around 60%.

Building Telco Reasoning Models for Autonomous Networks with NVIDIA NeMo

NVIDIA and Tech Mahindra have detailed a collaborative effort to build specialized AI reasoning models for telecommunications, aiming to accelerate the adoption of autonomous networks. The project addresses a persistent skills gap in the industry, where telcos prioritize AI-driven automation but often lack the internal expertise to implement it. By creating a repeatable pipeline using open models and tools, the companies are providing a blueprint for operators to develop AI agents that can automate high-volume tasks within their Network Operations Centers (NOCs).

The technical approach bypasses the complexity of training on raw network logs by using the NVIDIA NeMo toolkit to orchestrate a structured pipeline. The process begins with generating synthetic, yet realistic, incident data. Human expert procedures for fault management are then translated into structured reasoning traces, which serve as a curriculum for fine-tuning a Qwen3-32B large language model. This method teaches the model to emulate an experienced engineer's workflow for triage, root-cause analysis, and resolution. Evaluations of the resulting model demonstrate a significant increase in accuracy for key tasks, rising from a baseline of approximately 20% to 60%.

This development offers a practical path for network operators to transition from manual, reactive alarm-handling to closed-loop, automated systems. By deploying specialized agents to handle routine incidents, telcos can reduce mean time to resolution (MTTR) from hours to seconds for common faults, directly impacting operational expenditures. This allows human engineers to shift their focus from repetitive triage to proactive network optimization and more complex problem-solving, pushing the industry closer to achieving highly autonomous network capabilities, such as TM Forum Level 4.

The collaboration's core innovation is its reproducible pipeline that transforms human expertise into structured, synthetic training data. This strategy enables telcos to build specialized, tool-calling agents without requiring massive volumes of curated production logs, presenting a more scalable and accessible path to network automation.