TL;DR
- Open Model: NVIDIA released an open 30-billion-parameter Nemotron Large Telco Model at MWC Barcelona, built with AdaptKey AI to automate fault isolation and network remediation workflows.
- Accuracy Gains: The model triples incident summary accuracy from roughly 20% to 60%, reducing the volume of alerts that require human review in network operations centers.
- Global Deployments: Cassava Technologies, NTT DATA, and Telenor Group are already deploying NVIDIA’s agentic network configuration blueprints across Africa, Japan, and maritime 5G environments.
- Industry Access: NVIDIA released the model and an implementation guide through GSMA’s new Open Telco AI initiative, giving mid-tier carriers without dedicated AI teams access to autonomous network tooling.
NVIDIA last week handed telecom operators a free, open 30-billion-parameter AI model trained specifically on network data. Purpose-built to replace engineers who work overnight shifts isolating faults and manually rolling back failed configuration changes, the model debuted at MWC Barcelona alongside agentic blueprints already running on three continents.
Timed to Mobile World Congress Barcelona running March 2–5, 2026, the announcements center on an open Nemotron-based large telco model (LTM) alongside new NVIDIA Blueprints for intent-driven RAN energy efficiency and multi-agent network configuration. Jensen Huang set the stakes at MWC:
“AI is redefining computing and driving the largest infrastructure buildout in human history – and telecommunications is next. Together with a global coalition of industry leaders, NVIDIA is building AI-RAN to transform the world’s telecom networks into AI infrastructure everywhere.”
Jensen Huang, founder and CEO of NVIDIA (via NVIDIA)
The Open Nemotron Telco Reasoning Model
That ambition starts with the model itself. NVIDIA collaborated with AdaptKey AI to release the open source 30-billion-parameter NVIDIA Nemotron LTM for telecommunications operators. Optimized for telecom-specific terminology, the model reasons through workflows such as fault isolation, remediation planning, and change validation.
Despite strong operator interest, the tooling to act on it has been largely absent. The NVIDIA State of AI in Telecommunications report found 65% of operators said AI is driving network automation, with 50% naming autonomous networks as their top AI use case for return on investment. Yet without domain-specific models trained on telecom workflows, that priority remained aspirational.
How the Model Was Built
Built on the NVIDIA Nemotron 3 family of foundation models and fine-tuned by AdaptKey AI using open telecom datasets, including industry standards and synthetic operational logs, the LTM improves incident summary accuracy from roughly 20% to 60%.
That accuracy gain marks a practical threshold. Network operations centers processing thousands of alerts per day require engineers to filter noise and prioritize response. At 20%, every output demands human review; tripling that rate means the model now cuts the volume requiring human judgment, shifting the deployment calculus toward savings over oversight costs.
Agentic AI Blueprints in Operator Deployments
The reasoning model does not operate alone. VIAVI’s TeraVM AI RAN Scenario Generator powers the RAN energy efficiency blueprint, producing synthetic network data – cell utilization, user throughput, and traffic patterns – so operators can train and test agents without exposing live network traffic.
Moreover, early adopters already span three continents. Cassava Technologies is using the network configuration blueprint to build the Cassava Autonomous Network, deploying three coordinated agents to monitor, apply changes, and roll back if needed across Africa’s multi-vendor mobile environment. NTT DATA is deploying the blueprint for intelligent traffic regulation with a tier-1 Japanese operator.
Building on those deployments, Telenor Group will be the first telco to formally adopt the blueprint through BubbleRAN, targeting maritime 5G connectivity for ships operating at sea. This breadth validates the agentic architecture against meaningfully different constraints: multi-vendor interoperability, enterprise traffic management, and intermittent at-sea connectivity.
Open Ecosystem and Operator Access
Those early wins are only possible because NVIDIA deliberately lowered the barrier to entry. NVIDIA and Tech Mahindra published an open source guide showing operators how to fine-tune domain-specific reasoning models and build agents for network operations center (NOC) workflows. Released through GSMA as part of its new Open Telco AI initiative, the LTM, implementation guide, and blueprints extend access to operators without dedicated in-house AI teams.
However, the release is also a strategic move. Channeling the model through GSMA signals NVIDIA is targeting mid-tier carriers; operators that would not engage with NVIDIA’s commercial stack can now access the model through an industry body they already participate in.
Long-term, NVIDIA targets TM Forum Level 4 highly autonomous networks, where systems handle the majority of incidents without human approval. The GSMA Open Telco AI initiative provides the institutional runway to make that progression industry-wide.
For network engineers currently managing overnight fault shifts, the timeline is now visible. Carriers that begin fine-tuning the open LTM on their own infrastructure data will be positioned to reach Level 4 autonomy ahead of peers, while those that delay face compounding labor cost disadvantages as the automation gap widens.
