Meet the CLM: The Next Era of AI Is Not Just Another LLM

The first wave of generative AI gave us extraordinary generalists: ChatGPT, Claude, Gemini, and others that can write, reason, code, and work across an enormous range of subjects. But they are prone to hallucinations and drift. This is partly because their learning methodologies are vast and distanced. Instead of learning the techniques themselves, Large Language Models (LLMs) simply “read” how to do things. If they are not tightly constrained, they go haywire. 
 

This also concerns the training methodologies of LLMs. Once they become so large, they tend to drift back to their original training. So it does not matter if you keep pointing it at something new, it will ultimately return back to the first thing you trained it on. That is a big problem for accuracy and learning over time.  
 

This is why the industry pivoted to RAG architectures and knowledge bases. For a number of use cases, this is a perfectly suitable solution. But it is still just “reading,” only now pointed at a specific volume, so it is more contained. We simply constrained the architecture to a narrower set of relevant knowledge.  

The next wave is about “training experts” 

The next wave will be different. It will be built around natively-trained specialists: intelligence designed for a specific job, sized for that job, and placed directly where it can perform best. But these models are not learning where the knowledge resides. They are learning the expertise itself. It is a small and medium sized language models, but they have evolved to understand the knowledge natively. In the context of broadband and Calix, it is what we have invented: a communications language model (CLM).  
 

For communications service providers (CSPs), this shift is more significant than the early unleashing of LLMs upon the world. Network troubleshooting, configuration diagnosis, and threat detection are not the same job. You would not ask an orthopedic surgeon to repair your car. Or a mechanic to perform surgery. The right expertise changes the result. Our CLMs learn from the expertise we have contained within each model itself, not from the entire internet, and they have become experts in their field.   
 

We split the corpus of knowledge into manageable pieces and first trained a number of subject matter expert teachers. Then we tested the teachers to make sure each was indeed an accurate expert. Next, the teachers each serially train the CLM in each of their fields of expertise. What you are left with at the end is a model that is now natively an expert in each field and is as accurate as the teachers. They are also right-sized, to run at the network and subscriber edge on the existing appliances, not just in the cloud on GPUs and not dependent on a RAG based knowledge base.  
 

Our open-weight CLM, publicly available on Hugging Face, is a first step: shifting from constrained general-purpose intelligence into a team of communications specialists. It is an accurate, predictable 3-billion-parameter, medium-sized CLM that was quickly and cost-effectively trained using RSOD techniques.  

A CLM is not just another worker agent (it has never existed before) 

Specialization is not enough. A CLM is not the same as a worker agent.  
 

Beginning in 2027, more of these CLMs will appear. They are reproducible, cost-effective, and capable of retaining deep communications expertise from multiple specialized teacher models. This means we are no longer constrained to the cloud for many of our top distributed use cases like cybersecurity, repair, and troubleshooting.  

We will pair the CLMs with an agent and other components to perform specialized operational work locally and perform autonomous decision-making at the network and subscriber edges.   
 

Nothing like this has ever existed.  

Predictable learning means predictable cost 

CSPs do not need more AI experiments. They need AI that performs in the business. 
 

That means intelligence they can trust and economics they can predict, in both cost and return. 
 

Predictable cost means subscription economics, without variable token charges or overages. Predictable return means leveraging their entire asset as they build an AI-powered service provider: the network, the customer premise, and the cloud to deliver value. 

CLMs will absorb the complexity of AI engineering for service providers 

We judge technology by what it enables service providers to accomplish. 
 

Across Calix customers, Net Promoter Scores℠ (NPS®) have reached the 80s and 90s. Operating expenses have declined by as much as 80 percent. Incremental monthly recurring revenue has increased by 500 percent. Truck rolls have declined by more than 40 percent. Service adoption has increased fivefold. 
 

AI can accelerate outcomes like these. But only if we start with the outcome and measure what intelligence delivers. 
 

Service providers should not need teams of AI engineers continually evaluating models, architectures and infrastructure as the technology evolves. Our job is to absorb that complexity. 
 

The standard should be uncompromisingly clear: Know the cost. Define the outcome. Measure the result. 
 

That is how AI stops becoming another Gordian knot of models, agents, infrastructure, and siloed data. And starts becoming an operating advantage that helps every service provider compete and win. 

Net Promoter ScoreSM (NPS®)
Net Promoter®, NPS®, NPS Prism®, and the NPS-related emoticons are registered trademarks of Bain & Company, Inc., Satmetrix Systems, Inc., and Fred Reichheld. Net Promoter Score℠ and Net Promoter System℠ are service marks of Bain & Company, Inc., Satmetrix Systems, Inc., and Fred Reichheld.