Telefonica runs compressed Llama models on its own infrastructure for AI agent-assist for over 8,000 customer-service agents

Overview

Multiverse Computing built compact models (compressed and optimised Llama 3.1 8B and Llama 3.3 70B) that run entirely on Telefonica's own infrastructure, giving more than 8,000 customer-service agents AI-assisted support as part of the 'Movistar por ti' customer-care initiative. Keeping customer data on the operator's network ruled out external cloud APIs. The full story is gated.

The challenge

Telefonica wanted to scale generative AI agent-assist across thousands of contact-centre seats as part of its "Movistar por ti" customer-care initiative, but frontier large language models were too costly and resource-intensive at that scale. Keeping customer data on its own network ruled out external cloud APIs, and running full-sized models on-premises would have required infrastructure investment beyond what the rollout could justify, limiting it to small pilots.

The solution

Multiverse Computing built compact models, compressed and optimised versions of Llama 3.1 8B and Llama 3.3 70B, that run entirely on Telefonica's own infrastructure and give more than 8,000 customer-service agents faster, AI-assisted support. The rest of the story is gated and not visible.

Reported business value

The page headline says the work cuts infrastructure costs by 72%, but the visible text does not explain how that figure was measured; the full story is gated. The visible summary states that more than 8,000 customer-service agents are given AI-assisted support.