Telecommunications is an incredibly complex, highly specialized domain. Modern mobile networks are inherently multi-vendor, 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. When operating at an immense global scale that connects 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 specialized domain knowledge.
While general-purpose AI models have come a long way, the scale, complexity, and specificity faced by telecom providers means domain-specific models remain the best way to achieve the dramatic network and process automation and agentic workflows that are at the heart of the AI era. And it takes an open model to deliver the flexibility and dynamism global networks require.
Why domain-specific models matter
Generalized frontier models are incredibly capable at broad reasoning and language tasks, but they lack the foundational context required to manage critical infrastructure. General models still struggle with highly specialized 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 models: Underpinning Open Telco AI
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 Gemma models.
These models were trained on a specialized 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, optimizing the balance between accuracy and 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 that are so central to modern life.
“The Open Telco AI platform represents a critical milestone in establishing trusted, domain-specific intelligence for the telecommunications industry,” said Louis Powell, director for AI technologies at GSMA. “By leveraging open-source foundations like Gemma, we are proving that highly accurate, efficient, and reproducible models can be built through global industry collaboration.”
Gemma emerges as a leading model
AT&T’s tests during OTel development 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 returned correct response 91.74% of the time, achieving the highest overall accuracy for all models tested.
-
This baseline version of Gemma 3 with 27-billion parameters delivered the strongest performance in initial model training across the models tested by AT&T.
-
The Gemma 3 model with 300-million telco-related embeddings saw a significant retrieval improvement.
“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. This can help increase accuracy while driving down costs at the same time.”
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, demonstrating that tailored, smaller models can outperform massive frontier models when optimized 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-optimized 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.
