Generative AI, and particularly LLMs, are gaining business traction.  However, as organizations move from testing, experiments and proof of concepts to production deployments that can have material impacts on businesses, teams must consider how to do this in the right way for their business.

There are generally two options to use:

  1. (Very) large language models supplied by the big guys, running in their cloud infrastructure. These are huge models that – at least – claim to be able to take on most tasks.
  2. Smaller, efficient, self-trained language models that run on your own infrastructure.


It may not always the best option to go with the externally developed large language model route.  When going down the route of the second option, you still end up with a generative AI language model.  However rather than using someone else’s model, you fine tune or train you own language model, on your own data, targeted at your own requirements, and run this in your own environment.

Here’s why this approach can be beneficial:

Avoid vendor lock in.  Generative AI is an incredibly fast-moving space and you need the flexibility to change as your needs develop.  Should you be going all in with one model vendor and following their chosen path rather than yours?

Control your data.  To use an LLM you’re going to need to send customer, employee or other confidential data to it.  Ask yourself if you should (or are even legally allowed) to send this out to a third party, or if this should be kept internal.

Focus on your business tasks.  Huge LLMs take on a wide range of tasks.  However, do you need a language model that can write poetry and make jokes if you’re a bank or do you need it to address your own business tasks?

Cost & predictability.  Small models are more efficient to run, with lower hardware requirements.  Take the new AI PCs (or upcoming Copilot+ PCs in Microsoft language) – with an on device NPU do you need to spend big on running a huge model?

Integrate where you like.  It may be quick to connect to an external API in a small project.  But will this access always be available, or will the vendors pull up the drawbridge further down the line in attempts to control the market?

Clarity on the process.  Your business will have processes and controls for how you develop and use technology (which you’ve been using for years).  By following the model training process yourself, you can apply your controls rather than trusting the external vendor on this.

Once your business is in full control of your own small, dynamic language models new options can open up to you.  To date, use cases for generative AI have typically focussed on productivity gains alone.  Whilst this is a sensible first step, the benefits of small language models can empower your organization to move into new territory and start deploying revenue generating use cases.  This alone, would be a significant shift for the generative AI space.

As generative AI is flourishing into production business use cases, perhaps small language models are really the future..

Back to Blog