AT&T and GSMA’s top-ranked OTel 2.0 open-source model gives operators a high-performance alternative to proprietary, closed AI ecosystems.
On 24 July 2026, AT&T and the GSMA announced the release of OTel 2.0, an open-source, post-trained domain adaptation of Gemma 4 31B-IT. Purpose-built for the telecommunications sector, the model currently holds the number 1 spot on the Open-telco.ai benchmark leaderboard, establishing itself as the top-performing open-source AI model designed specifically for telecom operations.
By offering open model weights, the release directly tackles closed-ecosystem lock-in, addressing a major operational hurdle where general-purpose frontier LLMs struggle to parse dense network telemetry, standards documentation, and live system troubleshooting. The model’s foundation rests on a multi-vendor open-data ecosystem rather than a single proprietary pipeline. Post-trained on a 400 billion token telecom dataset filtered from over 1 trillion processed tokens, OTel 2.0 incorporates a 15-billion-token base provided by the GSMA, compiled from seven major standards development organisations (SDOs), including 3GPP, ETSI, GSMA, CAMARA, ITU, O-RAN, and TM Forum.
It also integrates with the Telco Corpus, an open 10-billion-token technical specification commons released jointly by the GSMA and Pleias on 25 June 2026. Developed alongside Red Hat, Dell, Microsoft Azure, and AMD, the processing and synthetic data pipelines leveraged open models (such as Microsoft’s Phi-4) on Azure compute, avoiding expensive closed-API endpoints. Unveiled by AT&T CTO Jeremy Legg during an AMD keynote, OTel 2.0 was trained directly on AMD Instinct hardware, highlighting a strategic move toward hardware and data sovereignty that deliberately avoids single-vendor lock-in to CUDA-dependent chipsets or proprietary cloud environments.
Building on the success of OTel 1.0—which recorded over 18 million downloads following its March 2026 launch—the new 31-billion-parameter architecture proves that smaller, domain-adapted open models can consistently outperform massive closed frontier LLMs on specialized tasks. This approach dramatically lowers deployment costs and energy overhead for hybrid multi-cloud and on-premise network environments.
