Open-weight model

An open-weight model is a language model whose trained parameters are published for download, letting an organization run it on its own infrastructure instead of calling a vendor's API. This is not the same as open source: training data and training code are often withheld, and the license may restrict commercial use or large-scale deployment. Open weights matter for private deployment because they let a team run LLM inference inside its own environment rather than sending data to a third party. Example: a healthcare provider selects a downloadable model so patient call transcripts never leave its network, supporting an air-gapped deployment where full internet access isn't an option.

An open-weight model is a language model whose trained parameters are published for download, letting an organization run it on its own infrastructure instead of calling a vendor's API. This is not the same as open source: training data and training code are often withheld, and the license may restrict commercial use or large-scale deployment. Open weights matter for private deployment because they let a team run LLM inference inside its own environment rather than sending data to a third party. Example: a healthcare provider selects a downloadable model so patient call transcripts never leave its network, supporting an air-gapped deployment where full internet access isn't an option.

Related terms

Large language model

A large language model is the model class behind modern conversational AI, summarisation, and reply drafting. Trained on broad text corpora, it predicts language fluently but holds no inherent knowledge of a specific business, which is why grounding and retrieval are required before answering customer questions.

LLM inference

LLM inference is the act of running an already-trained language model against new input to produce an output. This is distinct from training, the earlier process of building the model's underlying parameters. In a contact center, every AI-generated reply, conversation summary, or quality score is the product of one or more inference runs. Each turn in an automated chat exchange triggers at least one inference call, converting the customer's message and relevant context into a response. Inference is typically what usage-based billing or internal compute consumption is measured against, and it is the stage at which response speed is determined, since the model has already been built and is simply being queried. Input and output length for an inference call is counted in LLM tokens.

Air-gapped deployment

An air-gapped deployment runs software, including AI models, on infrastructure with no connection to the public internet, so data cannot leave the network boundary. Regulated industries such as banking, insurance, and healthcare request this isolation when compliance rules or internal policy prohibit customer data from touching any external link. The trade-off is that updates and model refreshes must be delivered manually instead of pulled automatically, and integrations with outside systems are constrained. This differs from on-premise or private cloud setups, which may still route traffic through encrypted internet connections for vendor management or updates. A contact center example: a bank running its support assistant entirely inside its own isolated network, so call transcripts never cross an external connection. OnClarity holds SOC 2 Type II certification and is GDPR and HIPAA Ready, with Saudi PDPL alignment, for enterprises evaluating strict deployment requirements.

On-premise

On premise means software runs on servers the customer owns or controls, inside their own data center or private cloud, instead of the vendor’s environment. In an AI contact center, on-premise deployment gives regulated operators direct control over where conversation data is processed and stored. This matters when telephony investments, internal audit rules, or sector regulation make external hosting difficult to justify. Example: a utility keeps call recordings inside its own network, while the AI model itself runs in an isolated environment the utility controls directly. The trade-off is real. The customer takes on hardware cost, patching and upgrade cycles, and the engineering staff needed to run the stack. The vendor no longer carries that work. Many enterprises land on a hybrid model instead. They pair on-premise components with private cloud for everything that does not require strict data isolation, which keeps costs lower than a fully on-premise build.

OnClarity

5,000 free AI support resolutions for startups

Startups in the US, UK and Europe with under 200 staff get 5,000 free resolutions from OnClarity’s AI support agent as their trial.

Apply for the Startup Program

Open until 31 March 2027 or until 50 startups are approved.

OnClarity

5,000 free AI support resolutions for startups

Startups in the US, UK and Europe with under 200 staff get 5,000 free resolutions from OnClarity’s AI support agent as their trial.

Apply for the Startup Program

Open until 31 March 2027 or until 50 startups are approved.

OnClarity

5,000 free AI support resolutions for startups

Startups in the US, UK and Europe with under 200 staff get 5,000 free resolutions from OnClarity’s AI support agent as their trial.

Apply for the Startup Program

Open until 31 March 2027 or until 50 startups are approved.