AT&T and GSMA build open telco AI on Gemma, hitting 91.7% accuracy

Telecom is a uniquely hard domain for AI because its specialized vocabulary, multi-vendor data structures, and proprietary telemetry are almost absent from the open internet. There is no public corpus for network operations. GSMA Intelligence reports that only 16% of AI deployments in telecoms are on the network, precisely because general models lack the foundational context needed for real-time infrastructure management. The article exposes this tension: frontier models are good at reasoning, but they cannot safely interpret network logs or diagnose performance bottlenecks without domain-specific tuning.

To solve this, GSMA launched the Open Telco AI platform, and AT&T post-trained a family of open telco models called OTel on architectures including Google’s open-source Gemma models. The initiative delivered 30 models of varying sizes, trained on a specialized telco dataset curated by operators, equipment vendors, and academia. Safety was built in from the start: models were trained for abstention using retrieval augmented generation (RAG) to reduce hallucinations, a critical requirement in regulated environments. AT&T‘s tests showed that Gemma-4-E4B-it achieved 91.74% accuracy, the highest among all tested models, and Gemma 3 with 300M telco-related embeddings saw significant retrieval improvement. AT&T‘s VP of data science and AI noted that domain-specific fine-tuning lets smaller models outperform legacy models several times their size, driving down costs while increasing precision.

For builders, the takeaway is concrete: open, domain-specific fine-tuning beats generalized scale in specialized operational settings. The OTel models have already been downloaded over 18 million times and rank at the top of the Open Telco Benchmarks. The lesson extends beyond telecom—any industry with sparse, high-stakes domain knowledge (networking, medical devices, industrial control) should consider building small, fine-tuned models with built-in abstention and RAG rather than waiting for frontier models to absorb their niche. Google Cloud’s full-stack approach—infrastructure, tools, and open models like Gemma—makes this path practical for operators to replicate on their own data.