Personal AI’s core model, MODEL-2, is a small language model. In this context, “small” refers to the number of parameters when compared with LLMs. Where a model like Meta Llama 3 can have up to 75 billion parameters, MODEL-2 is comparably smaller, more efficient and tuned for training on your data.

SLM vs LLM: more grounded on your data

The objective of a general LLM is to provide the best possible answer about any topic, in any kind of writing style. As a result, machine learning teams round up a massive amount of information, so the model can train and grow its experience. These models are meant to represent all expertise across any area. But this flexibility can be their greatest drawback: after all, hallucinations are still a risk, especially in topics where the model wasn’t as highly trained on.

An SLM’s lower flexibility when compared with an LLM is actually an advantage. You want your Personal AI to represent you, your company and your products and services. While that’s a lot of data, it’s not as much as the entire human knowledge, so having a model that’s more inflexible means that it will stick to your facts, not try to guess with its general training experience.

SLM vs LLM: more efficient

As AI models grow bigger, they need more powerful hardware to compute all the calculations both to train and to generate answers. This creates a context where all the major providers are rushing to build AI data centers to host their AI models. As the industry progresses, it’s unlikely you’ll be able to run a flagship model in hardware physically installed in your company premises—unless you’re willing to commit to a large upfront investment and high maintenance costs.

SLMs, since they’re smaller, will be able to run on less powerful hardware, someday even in mobile phones.