AT&T and GSMA Tap Gemma to Supercharge Telecom AI

Telecom networks are a tangled swamp of proprietary data and multi-vendor chaos. While AI has conquered coding and general language, the telecom domain remains a murky deep pool—there’s no Wikipedia for network logs. That’s why AT&T and GSMA are diving in with Google’s open-source Gemma models to build domain-specific AI that actually understands the swamp.

Why Domain-Specific Models Matter

General-purpose models are like caimans—big, impressive, but out of their depth in the telecom marsh. They choke on specialized vocabulary, complex topologies, and vendor-specific telemetry. Telco-specific models, trained on curated datasets, can interpret technical logs, diagnose bottlenecks, and follow industry protocols with the precision needed for real-time systems.

Open Telco AI: A Collaborative Swamp Drain

The GSMA launched the Open Telco AI platform to build trusted, telco-grade AI. AT&T post-trained a family of models called OTel on Google’s Gemma architectures. These models were trained on a specialized dataset curated by GSMA and collaborators, including operators, equipment vendors, and academia. The result: 30 models across various sizes, balancing accuracy and efficiency. Safety is built in via RAG to reduce hallucinations—critical in regulated telecom environments.

Gemma Shines in the Swamp

AT&T’s tests show Gemma outperforming other architectures after fine-tuning. Key stats:

• Gemma 4 E4B-it achieved 91.74% accuracy—the highest overall.
• Gemma 3 27B delivered the strongest baseline performance.
• Gemma 3 with 300M telco embeddings saw significant retrieval improvements.

“Gemma models have been setting the standard for open-source fine-tuning,” said Mark Austin, VP of data science and AI at AT&T. “By training on telco data, we outperform legacy models several times their size, increasing accuracy while driving down costs.”

What’s Next?

The OTel models have already been downloaded over 18 million times and rank top on the Open Telco Benchmarks. Google Cloud is providing a full-stack solution—AI-optimized infrastructure, development tools, and open models—so operators can fine-tune with their own data. The goal: replicate the AI progress seen in coding for telecom sub-domains like automated network configuration and self-healing systems. The swamp is getting smarter.

Original announcement published on Google Cloud.