Communications Language Models
Specialized AI built for service providers
General-purpose AI wasn’t designed for broadband. Communications language models (CLM) by Calix are purpose-built for communications service providers and power Calix Gems™, specialized operational AI that leverages CLM intelligence required to perform a particular task.
What is specialized?
“Right-sized” +“knowledge native
to the model” solving operational
problems at scale
A large, broadly trained model capable of reasoning across many domains. For specialized telecom work, Shane described the industry approach as constraining or augmenting the model with RAG, knowledge bases, or limited additional training. Local action is orchestrated and controlled centrally.
Small language models describes size, not specialization. More suitable to deployment outside the data center, yet a smaller generic model does not understand communications. Smaller models can be more amenable to the specialized training.
A small language model that has learned communications expertise natively. The CLM does not need an external knowledge corpus or require centralized orchestration to act. The CLM removes the need to carry a RAG architecture onto compute and memory constrained devices.
Calix white paper
Communications language models explained
An LLM is a general-purpose model that relies on external knowledge to become useful for specialized communications tasks. A CLM takes a fundamentally different approach: communications expertise is learned into the model itself. That makes the intelligence smaller, specialized and self-contained enough to move closer to where operational problems occur, creating the foundation for distributed AI across the cloud, network and subscriber premises.
Executive interview
Why Calix created CLMs
Michael Weening, President and Chief Executive Officer, and Shane Eleniak, Chief Product Officer, are featured speakers at the CNBC TEC Summit 2026, which runs from October 27–28, 2026. Their core commentary will emphasize how the AI industry has been racing to build bigger, general-purpose models that try to do everything, while the next phase is specialized: right-sized models with deep expertise in one industry and, over time, one job.
Why specialized models outperform general-purpose AI
CLMs are specialized AI models designed for communications environments. Calix Gems are purpose-built CLMs optimized for specific communications domains, delivering more expertise, greater efficiency, faster performance, and deployment flexibility than broad, general-purpose AI models.
Communications-specific expertise: Built and trained for targeted communications domains such as subscriber experience, marketing, security, and network operations.
Greater efficiency through specialization: Focused models require fewer compute resources while delivering more relevant and actionable outcomes.
Faster responses and insights: Purpose-built intelligence enables quicker processing, decision-making, and execution.
Optimized for every eeployment environment: Right-sized for their intended use case, from cloud platforms and network infrastructure to gateways, appliances, and subscriber-edge devices.
Calix AI Labs white paper
Calix releases universal communications language model on Hugging Face
The Calix Communications Language Model is now publicly available on Hugging Face. Learn more about the model and Calix’s approach to specialized communications AI in the accompanying technical paper: Building Telecom Language Models with RSOD: Domain Data, Specialist Post-Training, Memory-Conscious Distillation, and Networking Evaluation.
Communications Language Models and Calix GEMS
Frequently asked questions
Calix released a communications-domain AI model on Hugging Face, accompanied by a white paper explaining the approach and an evaluation framework and documentation on GitHub. Calix is also developing Calix Gems—proprietary specialists for Calix One customers, initially in three focus areas: security threat detection and mitigation, subscriber-experience optimization, and multi-dimensional network troubleshooting—which will not be included in the public release. Gems build on existing Agent Workforce capabilities and Calix’s preparation of the platform for AI.
The public Calix CLM is built on an eight-billion-parameter open-weight Llama model, fine-tuned on networking protocols and communications-domain knowledge, including access and edge networking, Linux, and related hardware and software information. Developers will be able to download, modify, and further fine-tune the model under the underlying Llama license.
Calix is not building giant general-purpose models from scratch. It is specializing in models for communications and right-sizing them for particular jobs and deployment locations.
Calix fine-tuned the public CLM on networking protocols and communications-domain information, including Linux and related platform, hardware, and software knowledge—not subscriber data.
Calix has developed a communications-domain model, publicly released on Hugging Face on October 27, accompanied by a white paper explaining its approach. Calix is also developing Gems: specialized models for Calix One customers, planned for introduction in 2027. Initial focus areas are real-time security threat detection and mitigation, locally optimized subscriber experiences, and dynamic multi-dimensional network troubleshooting.
Gems will put specialized expertise where it can do the most useful work, from subscriber equipment and the access network to the cloud, using the model architecture and deployment location best suited to each task.
Calix is developing a communications foundation specifically suited to further specialization and deployment across subscriber equipment, access infrastructure, and cloud environments. The Calix CLM provides the starting point for Calix Gems: models further trained to become experts at particular operational jobs, beginning with three focus areas: security threat detection and mitigation, subscriber-experience optimization, and multi-dimensional network troubleshooting.
The public CLM embodies the basic approach of creating smaller language models with embedded communications expertise. Together with the accompanying white paper, it gives developers and providers a foundation to explore, evaluate, and build on. Calix Gems contain proprietary expertise that is not intended to be open to public inspection.
Calix is developing Gems as specialists in discrete areas of a provider’s business, starting with three focus areas: security threat detection and mitigation, subscriber-experience optimization, and multi-dimensional network troubleshooting. For troubleshooting and repair, for example, Calix specialists will be enabled to diagnose problems and determine appropriate fixes, with the aim of shortening resolution times and reducing operational effort. For security, we are training specialists to recognize and respond to changing threats rather than rely solely on known attack signatures. Right-sizing and distributing those specialists where the work happens will make these capabilities practical at scale while supporting reliable, secure connectivity and subscriber loyalty.
The next phase of AI in communications should focus on specialization, not simply bigger or more powerful general-purpose models. Models trained for specific operational jobs—and sized and placed where the work happens—will resolve problems faster and use computing resources more efficiently. Distributing that expertise across subscriber equipment, the edge, and the cloud also will serve millions of endpoints without sending every problem back to a large, centralized model.
This approach will help service providers keep pace with rapid innovation while improving subscriber experiences, cultivating stronger loyalty, driving revenue growth, and lowering the cost to serve subscribers.
Calix sees language models as a way to train specialists that can diagnose problems and determine appropriate responses as conditions change, rather than rely solely on predefined workflows or known attack signatures. The intent is to bring that adaptability to troubleshooting, repair, and security—not necessarily to replace other AI or automation.
A specialist is only as good as its mentor. Calix brings decades of experience working with communications service providers—understanding how their networks operate, how their teams diagnose and resolve problems, and how operational decisions affect the subscriber experience. Our approach is to bring that accumulated industry knowledge into training specialists built for the work providers actually do.
Frontier-model companies bring broad AI capabilities, while security vendors and networking suppliers bring expertise in their respective domains. Calix’s opportunity is in combining specialized intelligence with a deep understanding of the provider’s business and operating environment. This will enable Calix to train specialists in what to look for, which questions to ask, and what constitutes an appropriate response in context.
Releasing the universal communications language model publicly gives developers and providers a way to explore our approach, evaluate the model, and experiment with further specialization, subject to its final license. We want to encourage broader exploration of AI trained for communications work and give technical teams something tangible to assess.
Calix is sharing a universal model, not the complete operational capability. Calix retains the proprietary specialization embodied in Gems and the work of bringing those specialists into Calix One’s agent-driven workflows. The public release lets others explore the starting point; Calix will apply its commercial expertise to develop specialists that perform useful work for providers.
Calix Gems and the Calix One platform
This builds on Calix One’s agent-driven workflows by adding specialized communications expertise where the work happens. Rather than relying on a local agent to send every problem to a larger, distant model for an answer, our approach is to make the smaller, local model the expert—trained to diagnose particular problems and determine appropriate responses itself.
This places expertise across subscriber equipment, the edge, and the cloud, matching the model and its location to the task. Calix One provides the framework for coordinating that expertise with other agents, tools, and human decisions. The goal is not simply more automation, but more capable local problem-solving without centralizing every decision.
Calix is developing Gems for Calix One customers, with models right-sized for particular tasks and deployment locations—from subscriber equipment and the access network to the cloud. Intended deployment locations include GigaSpire systems and access systems such as the E9-2. Each Gem will use the model architecture and location best suited to its task; not every capability will run in every location.
We intend to bring specialists into Calix One as part of the platform’s agent-driven workflows—not as standalone tools that providers must integrate individually. The platform provides the framework for connecting their expertise to operational context, coordinating work, and applying the provider’s rules for action and human oversight. The goal is to complement existing systems and security tools, not require providers to replace them. Support for particular third-party systems and equipment will depend on the integrations available through the platform.
Products and services
Related capabilities from Calix
Communications service providers can intelligently streamline their network operations and harmonize both subscriber and network data for a true white-glove subscriber experience.
Make every interaction with subscribers count with insights that enable you to deliver the experiences they want while growing brand value.
Transform broadband operations into a subscriber experience differentiator with insights, predictive analytics, and automation.
Simplify support and improve operational efficiency to deliver an exceptional subscriber experience with service excellence.