Calix, a US company that sells software and equipment to broadband providers, has released an AI model built for telecoms and networking. In a release on 27 October, Calix said it had “released a communications language model on Hugging Face”, the site where developers share AI models.
The model is open-weight, which means anyone can download it and run it on their own hardware. Calix also said that “in 2027, Calix plans to introduce Calix Gems”, a family of what it calls “right-sized”, specialised AI models. Those are plans, and none has been released yet. Calix says the Gems will be proprietary and will not be part of the public release.
What Calix released
The model is called Llama-Calix CLM 1.0, where CLM stands for Communications Language Model. Its Hugging Face page says it “starts from a pinned Meta-Llama-3.1-8B-Instruct snapshot”, meaning Calix took Meta’s freely available Llama model and trained it further on networking material. It is released under Meta’s Llama 3.1 licence, and there is also a compressed version for smaller devices.
Calix says it is meant for explaining network protocols, answering technical questions, diagnosing faults and working with router command lines. Training took about 56.1 GPU-hours on four NVIDIA RTX 6000 Ada graphics cards, using an estimated 4.63 kWh of electricity, according to the model card.
How big is it?
The released model has about 8 billion parameters. It is built on Meta’s 8-billion-parameter model, and the Hugging Face file data lists about 8.03 billion parameters. A Calix blog post by chief product officer Shane Eleniak first called it a 3-billion-parameter model. On 29 October Calix updated the post to say the public model has 8 billion parameters, and added that it is “also developing smaller, specialized models for future Gems”.
How Calix says it performs
Calix tested the model on 2,850 networking questions it held back from training, with answers scored out of 25 by another AI model, Anthropic’s Claude Sonnet 4.5. Meta’s original Llama model scored 12.87. Calix’s version scored 21.71 when given a prompt matched to each topic, and 20.00 with a general prompt.
In a comparison chart on the same page, it scored 20.08, behind Google’s Gemini 3.8 Flash on 23.48 and NVIDIA’s much larger Nemotron Super 120B on 20.96, but ahead of OpenAI’s GPT-OSS 120B on 19.44. These are Calix’s own tests, and the card says the exact versions of the rival models “should be pinned” before publication. A paper on arXiv, submitted on 20 October and co-written by Eleniak, describes the method and gives the same scores. It has not been peer-reviewed.
Calix says it built its own telecom-focused evaluation framework, Calix Telco-Bench, because general AI benchmarks may not capture networking knowledge and reasoning. It says the test measures performance on networking-specific technical tasks, and is not intended to show that the model can troubleshoot networks on its own or that it outperforms other models across all telecom tasks. The Telco-Bench documentation makes a similar point, saying its scores “are not ground-truth accuracy, pass-rate, or proof of operational correctness”.
Small specialist models, with caveats
Calix’s argument is that networks need many small expert models, not one giant one. “The future of AI isn’t based on one large generic language model in a data center trying to do everything equally well,” Eleniak said in the release. Jeff Heynen of research firm Dell’Oro Group, quoted by Calix, said operators’ networks “are, in effect, distributed data centers” with spare computing power at the edge.
That fits a wider shift towards models small enough to run locally, such as NVIDIA’s 4GB Nemotron 3.5 Bolt E4B. Telecom firms are also moving fast on AI agents; SK Telecom now gives its agents employee IDs. An open model trained on networking knowledge gives smaller broadband providers something they can test without sending data to a cloud AI company.
But the release raises questions. The benchmark was built and marked by Calix, using an AI judge, and has not been checked by anyone else. The model card itself warns the model “can produce incorrect, incomplete, outdated, or fabricated technical statements” and says letting its commands “execute unattended on production infrastructure” is outside the release’s intended scope “unless protected by additional validation, policy, and human-review controls”. It is also “primarily trained and evaluated in English”. Other open releases, such as Shanghai AI Lab’s Atria Dawn, show how crowded the open-weight field has become.
Calix says the public 8-billion-parameter model is open-weight and can be used for research or in production, subject to Meta’s Llama 3.1 Community License. That licence, shipped with the model, comes with conditions: anyone making the model or a product built on it available has to display “Built with Llama”, use has to follow Meta’s acceptable use policy, and companies that had more than 700 million monthly active users when Llama 3.1 came out need a separate licence from Meta. The model card adds that Calix’s own additions may carry further terms, lists research, development and evaluation as the intended uses, and says anyone using it in production should independently validate its commands and conclusions.
What comes next
Calix says the first Gems will focus on “real-time security threat detection and mitigation, locally optimized subscriber experiences, and dynamic multi-dimensional network troubleshooting”. Chief executive Michael Weening said: “In 2027, we will deliver more capabilities for customers than we did in the previous two years.”
Sources
- Calix via Business Wire: “Calix Releases Universal Communications Language Model on Hugging Face While Calix Gems Will Bring Right-Sized, Specialized AI to Network and Subscriber Endpoints” (27 October 2026)
- Calix blog, Shane Eleniak: “Meet the CLM: The Next Era of AI Is Not Just Another LLM” (26 October 2026)
- Calix AI Labs on Hugging Face: Llama-Calix-CLM-1.0 model card (read 27 October 2026)
- arXiv 2610.27993, Calix researchers’ paper on the training method (submitted 20 October 2026)
- Calix AI Labs on Hugging Face: Llama 3.1 Community License Agreement file shipped with Llama-Calix-CLM-1.0 (read 30 October 2026)
- Calix AI Labs on GitHub: Calix Telco-Bench evaluation framework (read 30 October 2026)
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- NVIDIA’s Nemotron 3.5 Bolt E4B Packs an Agent Model Into About 4GB
- SK Telecom Gives Its AI Agents Employee IDs and Human Supervisors
- Shanghai AI Lab Releases Atria Dawn Open Weights With 512K Context
- Akamai Launches 12-Model Inference Service and an AI Agent Sandbox
- India’s Telecom Regulator Proposes Responsible AI Rules for Operators
Update, 30 October 2026: An earlier version of this article reported that a Calix blog post described the model as having 3 billion parameters. Calix has since updated the post (29 October) to say the public model has 8 billion parameters, in line with its Hugging Face page.
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X