AT&T unveils telco open AI model

US telecommunications giant AT&T has launched OTel 2.0, a specialized open AI model designed specifically for the unique requirements of the telecom industry. Developed in collaboration with the GSMA and technology partners including Microsoft and Dell, the model aims to improve accuracy for network operations while significantly reducing infrastructure costs. This shift toward domain-specific AI represents a strategic move for the sector to move beyond general-purpose frontier models that often lack the specialized data needed for complex network troubleshooting and configuration.

AT&T’s OTel 2.0 is built upon Google DeepMind’s Gemma 4 31B-IT, an open multimodal model capable of processing text, images, and video sequences. To ensure industry relevance, the model was trained using 400 billion telecom-specific tokens curated from a pool of more than one trillion processed tokens. This specialized training allows the model to interpret industry standards and manage complex data sets more effectively than general-purpose alternatives. The project involves a broad coalition of industry leaders, including the GSMA, Microsoft, AMD, Dell, and Red Hat, highlighting a collaborative effort to standardize AI within the telecommunications ecosystem.

Alongside the new model, AT&T introduced an AI Gateway designed to address the cost challenge associated with large-scale AI deployment. According to AT&T Chief Data and AI Officer Andy Markus, the company processes an average of 45 billion AI tokens daily. The gateway utilizes a proprietary cache-aware router to intelligently match tasks to the most cost-effective model, ensuring that high-sophistication frontier models are only used when necessary. This system is already in production and has reportedly reduced AI inference costs by up to 90%, saving the company millions of dollars while maintaining performance levels.

The GSMA supports this move toward domain-specific AI, noting that general-purpose models often fail when tasked with troubleshooting live network issues because their training data barely touches the telecom domain. According to the association, OTel 2.0 currently sits at the top of the Open Telco AI leaderboard, demonstrating that domain-adapted models can be both more accurate and significantly smaller in size than general models. For the broader telecommunications sector, these developments are critical for meeting enterprise requirements, maintaining control across cloud and on-premise environments, and automating essential functions such as network configuration and product development.