A closed-source LLM is a large language model whose source code is not publicly available. It is owned, developed, and maintained by an organization that keeps details such as the model’s architecture and training data private. 

These models are capable of generating language that appears fluent and human-like, and are used in tools such as AI chat assistants, content creation platforms, and advanced search applications. Examples include GPT-4 by OpenAI, Gemini by Google, Claude by Anthropic, and models from Cohere.

Closed-source LLMs are part of the wider field of artificial intelligence, closely tied to an area called natural language processing, which focuses on enabling machines to work with human language in meaningful ways. A closed-source LLM strengthens this capability by offering more refined outputs and handling language-based tasks with a level of accuracy and fluency that supports commercial use.

To develop a closed-source LLM, a company trains it on a large body of data using high-performance computing and internally managed processes. The result is a model that performs well across many language tasks and is made available through interfaces such as web platforms or APIs. The methods used to train and structure the model are not shared publicly, and control over its use remains with the organization that built it.

How do closed-source LLMs work?

As with other large language models, closed-source LLMs are trained to generate fluent and human-like text by learning patterns from large datasets.

They are developed and managed in a unique way: the training methods, data sources, model weights, and architecture are not made publicly available.

Moving from pre-training to release, integration and monitoring involves the following process:

Pre-training on private datasets

Companies train their closed-source models using extensive and high-quality datasets, which may include licensed content, publicly available data, and proprietary sources; however, the exact dataset compositions are typically undisclosed.

While the exact sources are not disclosed, the data is chosen based on the use cases the model is designed to support. Examples include customer service, technical support, or business writing.

During training, significant computing power is required. Companies will typically have custom infrastructure in place or use partnerships with cloud providers. Training enables the model to learn general knowledge and language structure, and it will also develop its ability to reason.

Internal release and control

Once the model is trained, it can be released through controlled environments such as a web interface or enterprise software. Users can interact with the model, but they won’t be able to examine how it was built or trained.

The model architecture — or how the neural network is structured — along with the training weights, which contain what the model has learned, are kept private. The code used for inference, which allows the model to generate responses, is also not publicly available.

Deployment and integration

After internal release, companies can access the model through secure APIs or vendor-managed software platforms. The model remains hosted by the original provider, and businesses can connect it to their own systems using API integrations or pre-built software connectors.

Once integrated, companies can use the model for their websites, chatbots, office tools, or internal systems without needing to manage the underlying infrastructure. The provider manages system performance and scales capacity when needed. They also ensure the model complies with data protection and security standards. 

Customization and monitoring

Once the model is deployed and integrated, providers will monitor its use. They may adjust responses over time by fine-tuning internally. Another action they can take is rolling out system updates based on usage patterns.

As for users, they will not be involved in fine-tuning. While some platforms can be lightly customized, such as offering changes to memory settings or prompt templates, the underlying model will not undergo any change.

Examples of closed-source LLMs

Many commercial AI tools are powered by closed-source LLMs, and they will operate in specific industries for a tailored set of tasks. Their names are typically more widely known than open-source LLMs. 

Choosing the right closed-source LLM depends on the use case. For example, general-purpose models are designed to perform a broad range of tasks, including conversation, summarization, content creation, and reasoning. 

GPT-4 by OpenAI powers tools like ChatGPT and Microsoft Copilot. These general-purpose models can generate emails, answer questions, and perform other tasks. Similarly, Claude by Anthropic is often used in enterprise chatbots and assistant tools.

For enterprise-focused reasoning and retrieval, companies may seek information retrieval and structured outputs that meet specific standards. Watsonx LLMs by IBM are tailored for regulated industries such as healthcare and finance. 

Closed-source LLM use cases  

Closed-source LLMs are used in many everyday tools and services. Despite being less transparent, they are valued for stability and ease of integration, especially in commercial settings.

Closed-source LLMs are beneficial for organizations where employees are apprehensive about using AI. McKinsey research reveals that 41% of workers express concerns about cybersecurity and the quality of AI outputs, indicating they may need additional support during AI rollouts. As a result, the accuracy and security benefits of closed-source LLMs are ideal for such a scenario.

Here are some examples of how closed-source LLMs are applied in practice:

  • Customer support automation: Many businesses use closed-source LLMs to handle customer service interactions. Claude by Anthropic is one example used to respond to questions and offer guidance. Support teams often rely on the model to manage routine inquiries so staff can focus on more complex situations.
  • Workplace productivity tools: Closed-source models also feature in tools that help people complete writing and data tasks. GPT-4 is used within Microsoft Copilot, where it performs activities such as suggesting wording changes and supporting users with spreadsheet explanations. The model works directly within the software, reducing the need to switch between tools.
  • Industry-specific applications: Large organizations in the healthcare, finance, and legal fields use LLMs through platforms like IBM Watsonx.ai. The models are adapted to handle sensitive information, generate reports, assist with reviewing long documents, and more. Companies choose these tools for their ability to meet regulatory and compliance needs.