For the last few years, artificial intelligence has seemed obsessed with one idea: bigger is better.
Bigger models. Bigger datasets. Bigger training runs. Bigger headlines.
But biology may be one of the first fields to seriously challenge that belief.
In genomics, researchers are trying to decode one of the most complex information systems in existence: the language of DNA and RNA. That means identifying patterns across massive sequences, predicting gene function, understanding regulation, and uncovering signals that are often hidden in places scientists still do not fully understand. It is exactly the kind of problem that seems like it would require the biggest, most powerful AI systems possible. Yet recent work suggests something surprising: smaller language models may be able to do far more in biology than many people expected.
That possibility is fascinating, because it changes the question entirely. Instead of asking whether we can build the biggest model for genomics, researchers are increasingly asking whether we can build the smartest one.
Biology is not just another AI benchmark
At first glance, genomics might look like a natural extension of language modeling. DNA and RNA are, after all, sequences. They have patterns, structure, repetition, and context. That makes them a tempting target for the same kinds of models that have transformed natural language processing.
But biology is not just text in a different alphabet.
The genome contains layers of meaning that are far more difficult to interpret than a sentence on a page. Only a small fraction of the human genome directly codes for proteins, while vast noncoding regions help regulate when genes turn on, where they turn on, and how strongly they are expressed. Those regions are not random filler. They are part of a complex biological system that scientists are still actively trying to decode.
That is why AI in genomics matters so much. If models can learn the hidden grammar of biological sequences, they could help researchers predict the effects of mutations, identify functional elements, improve disease understanding, and accelerate discoveries across medicine and biotechnology.
Why the field assumed bigger would win
For a while, the logic seemed obvious. In mainstream AI, larger models often produce better results. So if genomics is difficult, then larger genomic models should be the answer too.
And to be fair, large genomic language models have made real progress. Researchers have applied them to tasks such as regulatory sequence prediction, functional annotation, splice-site analysis, and RNA-related classification problems. These systems have helped prove that biological sequences can, in fact, be modeled in powerful new ways.
But as the field has matured, a more nuanced reality has emerged. In biology, performance does not come from parameter count alone. It depends on the data, the architecture, the training objective, the benchmark design, the downstream task, and whether the model is actually capturing biologically meaningful patterns instead of just optimizing a narrow metric.
That is where small language models start to become really interesting.
The case for smaller models
A recent preprint described small language model agents for genomics, exploring whether more compact systems could still perform strongly when paired with the right tools and workflows. Rather than relying only on raw scale, the framework used agent-style reasoning and external resources to solve genomics tasks. According to the paper, its best setup reached very strong GeneTuring benchmark performance while remaining dramatically smaller than the largest frontier models. Because this work is still a preprint, it should be seen as promising rather than definitive, but it highlights an important shift in thinking: capability in biology may come not just from size, but from design.
That distinction matters.
A smaller model does not necessarily need to “know” everything inside its parameters if it can retrieve the right information, break a problem into parts, and use specialized tools effectively. In a field like genomics, that may actually be a more realistic path forward than endlessly scaling compute.
In other words, the future of biological AI may not belong only to giant monolithic models. It may belong to systems that are compact, targeted, and strategically connected to scientific tools.
Why this matters beyond accuracy
This is not just a technical debate. It is a scientific and practical one.
If cutting-edge genomic AI only works at the scale of enormous compute budgets, then access will remain limited to a small number of elite labs and major companies. But if smaller models can perform well on meaningful tasks, the technology becomes much more accessible. More labs can experiment with it. More researchers can fine-tune it. More hospitals, startups, and students can actually use it.
That changes who gets to participate in the future of computational biology.
Smaller models are also often cheaper to train, easier to run, and faster to deploy. In a real research environment, those advantages matter just as much as leaderboard performance. A model that is slightly less glamorous but far more usable can end up having much greater real-world impact.
The caution scientists still need
Of course, this does not mean small models have already solved genomics.
The field still faces serious challenges in benchmarking, reproducibility, and evaluation. A recent Nature Methods feature on genomic language model benchmarking highlighted just how difficult it can be to compare systems fairly, especially when code, datasets, and evaluation pipelines are inconsistent or incomplete. In some cases, researchers found that genomic language models underperformed strong supervised baselines on carefully chosen tasks, which is a useful reminder that hype and scientific value are not always the same thing.
That caution is healthy.
Biology is too complex for simplistic narratives about one model type “winning.” Large models may still be incredibly useful for broad representation learning and foundation-scale pretraining. Smaller models may shine in specialized workflows, task-specific assistants, or efficient deployment settings. The real future is likely to be more layered than the usual bigger-versus-smaller debate suggests.
A different definition of progress
What I find most exciting about this shift is that it forces AI to adapt to science, instead of forcing science to adapt to AI hype.
For years, progress in artificial intelligence has often been measured in size: more parameters, more data, more power. But genomics may require a better definition. In biology, the most valuable model may not be the biggest one. It may be the one that is efficient enough to use widely, precise enough to support real research, and aligned enough with biology to produce insights scientists can actually trust.
That is a much harder standard.
And honestly, it is a much more meaningful one.
If small language models can succeed in genomics, even in carefully designed parts of the problem, they could help make biological AI more scalable, more democratic, and more useful. That would not just be a win for machine learning. It would be a win for science itself.
Maybe the future of AI in biology will not be defined by the biggest systems in the room.
Maybe it will be defined by the models that are smart enough to do more with less.
