Calix releases downloadable telecom AI model, keeping planned Gems proprietary

Calix released a downloadable AI model for communications networks on October 27, 2026, giving developers a telecom-focused alternative to starting with a general-purpose model. The release puts model weights on Hugging Face, but not the company’s planned operational specialists: Calix Gems remain proprietary and are scheduled for Calix One customers in 2027.

The public Communications Language Model is built on an eight-billion-parameter open-weight Llama model. Calix fine-tuned it on networking protocols and communications knowledge, including access and edge networks, Linux, and related hardware and software information. It is specializing an existing foundation rather than building a giant general-purpose model from scratch.

The technical distinction Calix draws is where the specialist knowledge lives. Its model page contrasts communications knowledge learned during training with a broadly trained model that consults an outside knowledge collection. That outside-information approach is commonly described on the page as RAG. Calix’s pitch is to put the expertise into the model itself.

Calix says this design removes the need to carry a separate retrieval architecture onto devices with limited computing power and memory. It also says its communications models do not require centralized orchestration to act. Those are the company’s architectural claims, underpinning its plan to distribute specialized AI across cloud systems, networks and subscriber premises.

Size alone is not the distinction. Calix explicitly separates a small language model from a specialized one: a smaller generic model still needs communications training. Its stated goal is to match both the expertise and the model’s size to a particular job and deployment location.

The future of AI isn’t based on one large generic language model in a data center trying to do everything equally well. It’s groups of specialized models dispersed across the cloud, networks, homes, and businesses.

Shane Eleniak, chief product officer at Calix

For developers, Calix says the public model can be downloaded, modified and further fine-tuned under the underlying Llama license. The release therefore offers a foundation that developers can adapt, with those permissions governed by that license. Calix also published an accompanying white paper and says an evaluation framework and documentation are available on GitHub.

That public foundation is distinct from the products Calix intends to sell through its platform. Its model page describes each Gem as combining communications-model expertise with agent capabilities for a specific operational task. Gems are being developed for Calix One customers and are not included in the public release.

Calix identifies three initial areas for Gems. The intended functions go beyond answering networking questions: the company describes systems that would help its agents optimize service, diagnose problems and respond to security threats. These are planned applications, not capabilities delivered to customers by the public model release.

• Subscriber experience: Calix says Gems would learn how each location uses the network, helping Calix One agents anticipate demand and resolve issues before subscribers notice them.
• Network troubleshooting: the planned specialists would connect insights from subscriber premises, access systems and network services so agents could identify root causes and automatically resolve complex problems.
• Security: Calix describes Gems learning normal device behavior and detecting unusual activity in real time, enabling agents to isolate threats before they spread and affect subscribers.

The deployment plan is not tied to one model architecture or one place to run it. Calix says additional Gems will use the architecture and location best suited to each task, spanning subscriber equipment, access networks and the cloud. Its announced introduction window is 2027; that roadmap remains a plan rather than current availability.