Small but mighty: could specialised AI models loosen Big Tech's grip?
A recent Financial Times article explored whether cheap, specialised AI models could threaten the dominance of Big Tech's frontier offerings. For years, domain-specific AI models, fine-tuned for fields like medicine, law, and finance, have promised to rival those of Big Tech organisations at a fraction of the cost. However, these models are often quickly overtaken by the next generation of larger, more capable ones. But recent developments suggest that fine-tuning at the level of individual firms, rather than entire professional domains, could provide impressive and potentially durable advantages.
The FT highlights two examples. The Chinese Kimi 2.6 model, which is an open-weight AI model and has one trillion parameters, was fine-tuned for the legal domain. Fine-tuning involves adjusting model weights, so having an open-weight AI model to start with is beneficial. According to the FT article, the results match those of a frontier model performance at an eleventh of the cost. However, given that training a frontier model is very expensive, we assume that an eleventh of that cost is still high.
In another example, a well-known hedge fund trained a model on the workflows of its own investment professionals, achieving a reduction in errors of almost 30 per cent compared to frontier models, allegedly at a fourteenth of the cost. Crucially, as these gains came from proprietary information and professional judgement inaccessible to frontier models, this advantage may prove more lasting than those seen in earlier experiments. Another alternative is to use retrieval augmented generation (RAG), which involves using proprietary information to supplement model prompts rather than adjusting model weights. While RAG is effective at smaller scales, it is more of an “add on” solution than a way of fundamentally changing model weights to improve generalisation ability in a particular problem domain.
The implications for Big Tech are significant. Does humanity want a world in which every company across every sector is ceding value to a few models that eat everything they see? A world in which businesses fine-tune smaller models on their own institutional knowledge would be less dependent on a handful of proprietary black-box providers, which many would see as a healthier market dynamic.
But this shift raises its own set of questions. Some of these experiments used Chinese base models, such as Qwen3-235B and Kimi 2.6. This could potentially expose businesses to a new type of geopolitical risk. Then there is the workforce dimension to consider: the hedge fund example, for instance, relied heavily on the tacit knowledge of its own investment professionals to train the model. Will employees willingly help to fine-tune AI tools based on the expertise they have spent decades building? Particularly if they fear that this knowledge could ultimately be used to replace them? The answer will likely depend on the balance of trust and power in each workplace, and it is a question that employers would be wise to consider carefully.
