- Calix released its communications language model (CLM) on Hugging Face and plans to begin introducing specialized models called Calix Gems in 2027
- The smaller models are designed for specific jobs including security, network troubleshooting and subscriber experience
- Calix says distributing those models closer to endpoints lets providers use existing compute while reducing the need to send every AI task back to a large cloud model
Calix is taking a different approach to telecom AI. Instead of trying to make one massive model smart enough to run everything, it wants to scatter smaller specialists throughout the network.
Today, the company released its communications language model (CLM) on Hugging Face and said that in 2027 it plans to introduce what it calls Calix Gems, or specialized versions of its CLM designed to perform specific jobs at different points in the network, from subscriber and network endpoints to the cloud. Initial areas of focus include security threat detection and mitigation, subscriber experiences, and network troubleshooting.
The idea is essentially to give a broadband network a collection of AI specialists rather than one giant AI generalist.
"The future of AI isn’t based on one large generic language model in a data center trying to do everything equally well," Shane Eleniak, Calix chief product officer, said in the company's announcement. "It’s groups of specialized models dispersed across the cloud, networks, homes and businesses."
In an interview with Fierce Network, Eleniak noted that large language models still make sense for tasks that span multiple domains, but plenty of networking problems are narrow, local and increasingly need to be handled autonomously. These include security, repair and troubleshooting.
Rather than using models with 30 billion or more parameters for those jobs, Eleniak said that smaller models can run on network-edge or premises compute, which changes the economics of deploying AI across a network.
"How do you best manage a network where you've got ... 10,000 [central offices]? You deploy 10,000 edge language models that use the memory and the compute that already exists there," Eleniak said. "So you're not paying for incremental compute. You're not paying for incremental memory. You're not paying for incremental power."
Small models, specialized jobs
Eleniak compared Calix's CLM and Gems strategy to how companies hire people. "We don't hire a security analyst and say, by the way, you need to do marketing and you need to do finance," he told Fierce. "We train them to be that specialist."
A specialist model trained deeply for security could be distinct from one designed for troubleshooting and repair, optimization or customer support. Because the models are smaller, Calix's vision is that multiple specialists could operate across the network rather than forcing tasks through one centralized model.
"We want to make them narrow and deep versus broad and narrow," Eleniak said.
That distributed architecture matters because broadband networks present both an AI computing problem and a networking problem.
For some tasks, operators need broad, centralized information. For others, the relevant information exists at an individual network location or subscriber endpoint. Eleniak said trying to bring information from millions of endpoints back to a "large brain" handling multiple other jobs doesn't make economic sense and also creates challenges because of the underlying physics of the network.
AI where the problem happens
Michael Weening, Calix president and CEO, said the approach is a choice about where AI should physically run.
"If you have a free CPU, why wouldn't you use it?" Weening said. He added that some applications may need to sit at the edge because they require a faster response, while others belong in the data center.
The vendor isn't abandoning bigger cloud-based models. Instead, its Gems strategy amounts to adding another tier to its AI architecture, he added.
Customers are already using agentic AI through Calix Agent Workforce, which is supported by the company's Google Cloud partnership. As Calix develops Gems, Weening said it will select the model architecture and deployment location according to the task and use case.
The broader bet is that the next phase of telecom AI won't simply be about adding increasingly large models. It will also be about figuring out how little AI an operator actually needs to solve a particular problem, and how close to that problem the intelligence can run.
A shift for Calix
Weening said Calix began reworking its platform for this shift in November 2023.
"Communications language models and Calix Gems are the next step, helping service providers harness specialized AI while continuing to drive the business outcomes that enable them to compete and win," he said.
Read more about Calix on Fierce Network:
Why Calix came to a cable show: Fiber gets customers connected. AI keeps them.
Data center labor shortage pushes operators to train, poach and automate
Calix bets agentic AI can offset broadband's aging workforce