The Specialised Challenge of Telecom AI
Telecommunications is an incredibly complex, highly specialised domain. Modern mobile networks are inherently multivendor, featuring diverse and often proprietary data structures. While AI has made massive leaps in general language and coding, telecom domain knowledge is rarely accessible on the open internet—there is simply no “Wikipedia” for telecoms.
This data scarcity creates a major hurdle for AI models trying to deeply understand network operations. Operating at an immense global scale, connecting billions of people hundreds of billions of times a day, the industry requires absolute precision. Yet, according to GSMA Intelligence, only 16% of total AI deployments in telecoms are on the network, largely due to the difficulty of training models on specialised domain knowledge.
Why Domain-Specific Models Matter
While today’s general frontier models are incredibly capable at broad reasoning and language tasks, they lack the foundational context required to manage critical infrastructure. General models struggle with the highly specialised vocabulary, complex network topologies, and vendor-specific telemetry data unique to the telecom sector. Telco-specific models solve this by anchoring the AI in the actual realities of network operations. By training on domain-specific datasets, these tailored models can interpret nuanced technical logs, diagnose network performance bottlenecks, and understand standard industry protocols with the high degree of accuracy and precision required for real-time systems.
Google’s Gemma Model: Driving the Open Telco AI Initiative
To address this challenge, the GSMA recently launched the Open Telco AI platform to build accurate, efficient, and trusted telco-grade AI. As a core part of this collaborative effort, AT&T post-trained a family of open telco models, called OTel, on different architectures including Google’s open source models called Gemma.
These models were trained on a specialised telco-specific dataset curated by GSMA and its collaborators, including telecom operators, network equipment providers, and academia. The initiative successfully delivered 30 models across a range of sizes and architectures, optimising the balance between accuracy versus efficiency.
Crucially, these models are built with safety at their core, being trained for abstention using retrieval augmented generation (RAG) to drastically reduce hallucinations—an absolute necessity in highly regulated telecom environments.
Key Findings: Gemma Emerging as a Leader
The findings highlight the strength of Gemma compared to other architectures, demonstrating strong performance gains across the entire OTel model family after telecom-specific fine-tuning. Notably:
- The gemma-4-E4B-it model has currently achieved the highest overall accuracy
- The 27B parameter baseline Gemma 3 delivered the strongest performance in initial model training
- The Gemma 3 300M embedding baseline saw a significant retrieval improvement
Empowering the Future with Google Cloud’s Full-Stack Solutions
The impact of this open collaboration has been immediate, with over 18 million downloads of the models to date. Today, OTel stands as one of the top Models on the Open Telco Benchmarks, proving that tailored, smaller models can outperform massive frontier models when optimised for specific domains.
Looking ahead, Google Cloud is committed to supporting telecom operators globally in developing and deploying their own custom telco AI models. By providing a comprehensive, full-stack solution—including robust AI-optimised infrastructure, AI development tools, and open models like Gemma—we can help operators, vendors, and innovators fine-tune these models further with their own data. This enables telecom operators to accelerate their journey in AI adoption while deploying telco-grade AI safely using Gemma’s built-in support and guardrails.
Together, the telecom industry can replicate the incredible progress seen in coding and reasoning, bringing those advanced capabilities into critical telecom sub-domains such as automated network configuration and self-healing systems.
Industry Perspectives on Collaborative Innovation
“The Open Telco AI platform represents a critical milestone in establishing trusted, domain-specific intelligence for the telecommunications industry,” said Louis Powell, Director – AI Technologies at the GSMA. “By leveraging open-source foundations like Google Cloud’s Gemma, we are proving that highly accurate, efficient, and reproducible models can be built through global industry collaboration.”
“Gemma models have increasingly been setting the standard for open-source fine-tuning,” said Mark Austin, VP of Data Science and AI at AT&T. “By training these models specifically on telco data, we’ll be able to outperform legacy models several times its size in certain telco scenarios, helping increase accuracy while driving down costs at the same time.“
“Google Cloud is committed to supporting telecom operators globally by providing a comprehensive AI stack to accelerate adoption,” said Sridhar Gollapudi, Global Telco Market Lead at Google Cloud. “By combining our robust, AI-optimised infrastructure and AI development tools with powerful open models like Gemma, we’re committed to helping operators and vendors fine-tune and build Telco-Grade AI safely and effectively.”
Read Google’s full article to learn how Gemma is advancing Open Telco AI.