All guides / AI explained
AI explained · Seneca guides
AI models explained: choose the right model for the job
The biggest model is not automatically the best choice for every job. A useful AI service matches the model to the work: fast and economical for routine tasks, with stronger reasoning available when the problem demands it.
A model is the engine; an application is the experience
An AI model is a trained system that turns inputs into outputs. A language model works with language; a multimodal model can also work with supported inputs such as images or audio. The chat interface, document search, permissions and links to business systems are separate parts of the application around that engine.
That distinction matters when buying AI. Two products can use the same underlying model but deliver very different results because they provide different information, workflows and controls. Choosing a model is one part of building a service people can use productively.
Large, small and reasoning models
Large language models can support broad tasks such as drafting, analysis and coding. Smaller models can be effective for narrower work such as sorting enquiries, extracting fields or classifying documents, with lower hardware requirements. Reasoning models devote additional processing to working through demanding problems, often trading speed and cost for stronger performance on those tasks.
For example, a service desk can use a compact model to categorise incoming requests and send complex technical questions to a stronger model. A document team might use a vision-capable model to interpret scanned pages before passing extracted text into a separate workflow. Match the capability to the task instead of paying for the most demanding option on every request.
Open weights, open source and hosted models
Open-weight models make their trained parameters available under a licence, allowing suitable deployments on your own infrastructure. The licence still matters: available weights do not automatically mean unrestricted commercial use or that every part of training is open source. Hosted models are accessed through a provider’s service; their processing takes place in that provider’s environment.
Model families and releases change quickly. Build around the capability, deployment rights and operational fit you need, with room to change the engine later. Seneca supports private models and selected hosted integrations through a managed environment, giving your business flexibility over data location, performance and supplier dependence.
Choose on useful work, not a leaderboard
Compare models using representative tasks from your own operation: an enquiry to answer, a document to extract or a report to prepare. Look at accuracy, time to a useful answer, required corrections, simultaneous users and total running cost. A model that gives a cheaper first answer can become expensive if staff repeatedly repair it.
Hardware matters for private deployment. Model size, context length and concurrent requests all affect memory and processing requirements. Quantisation can reduce memory use by storing model values at lower precision, with a possible effect on quality. We assess your workload and help build the business case for a focused deployment or a wider rollout.
A few useful terms
- Parameters
- The learned numerical values inside a model. Parameter count alone does not tell you how well it will perform your task.
- Multimodal
- Able to work with more than one type of information, such as text and images.
- Latency
- The delay before a response begins or a task finishes.
- Quantisation
- Representing model values with fewer bits to reduce memory and processing demands.