The genomics foundation model race has mostly been a human health story. Companies like Google DeepMind with AlphaFold and startups like Evo from the Arc Institute have captured attention by decoding protein structures and microbial genomes. Agriculture, meanwhile, has remained a quiet corner of the AI genomics landscape, relying on older statistical methods and conventional marker assisted selection that can take a decade or more to move a trait from discovery to field deployment. That changed this week when Living Models released BOTANIC, a family of transformer models trained specifically on plant DNA, and the implications stretch well beyond the lab.
What makes BOTANIC notable is not just what it does but the constraints under which it was built. The model was trained on roughly 1,600 curated plant genomes covering 43 species using only eight NVIDIA H100 GPUs. For context, many large language models in the text domain require thousands of GPUs and months of training. The compute efficiency here suggests that domain specific genomics models may not need the same brute force infrastructure that general purpose AI demands, a finding that matters enormously for research institutions and agricultural companies operating on tighter budgets than Big Tech.
The core architecture treats DNA sequences as language, which is not a new idea in computational biology. Nucleotide transformer models from groups like InstaDeep and DNABERT from various academic labs have explored this approach. But BOTANIC narrows the focus to plants, and that specialization appears to pay off. The model learns gene regulation patterns and trait associations specific to plant biology, then uses those learned representations to score genomic variants linked to climate adaptive traits like drought tolerance and heat resistance. Across 22 genomics benchmarks, it achieves state of the art results, which in a field where incremental improvements often take years of wet lab validation is a strong signal.
Why plants, and why now
The timing is not accidental. Global crop yields are under escalating pressure from climate volatility. The 2023 and 2024 growing seasons saw record heat events across major agricultural regions in South Asia, the American Midwest, and sub Saharan Africa. Traditional breeding programs, even those augmented with genomic selection tools, struggle to keep pace because they are fundamentally bottlenecked by biological cycle times. You still need to grow the plant, observe the trait, and iterate. What a model like BOTANIC promises is the ability to prioritize which crosses and which variants to pursue before seeds go into the ground, compressing the discovery phase dramatically.
This mirrors what happened in drug discovery when AI models began predicting molecular binding affinities. The molecules still need to be synthesized and tested, but the search space shrinks by orders of magnitude. For plant breeding, the equivalent leap would be moving from screening millions of potential crosses to focusing on hundreds, with higher confidence that the resulting lines carry the desired climate resilience traits.
The competitive landscape is thinner than you might expect
Despite the obvious importance of agricultural genomics, very few organizations have built foundation models for plant DNA. Most crop AI efforts have focused on computer vision for field monitoring, yield prediction from satellite data, or supply chain optimization. The genomics layer has been underserved partly because plant genomes are notoriously complex. Many crop species are polyploid, meaning they carry multiple copies of their entire genome, which creates challenges for sequence alignment and variant calling that human genomics researchers rarely encounter. Wheat, for example, has a genome roughly five times larger than the human genome with extensive repetitive regions.
Living Models chose to work across 43 species rather than focusing on a single crop, which is a deliberate architectural bet. A multi species model can potentially learn conserved regulatory grammar that transfers across plant families, much like multilingual language models learn shared syntactic structures. Whether that transfer learning holds up in practice for specific breeding applications is an open question, but the benchmark performance suggests the approach has merit.
What this means for the breeding industry
The major seed and agrochemical companies, Bayer Crop Science, Corteva Agriscience, Syngenta, and BASF, all have internal genomics programs. They have spent decades building proprietary germplasm databases and phenotyping infrastructure. A publicly available foundation model that performs well on plant genomics benchmarks creates an interesting dynamic. Smaller breeding companies and public sector research institutions in developing countries, which lack the resources to build these models internally, could gain access to capabilities that were previously the exclusive domain of multinational corporations.
This democratization angle is significant because many of the crops most vulnerable to climate change are staples in regions with the least breeding infrastructure. Sorghum, millet, cassava, and cowpea receive a fraction of the research investment directed at corn, soy, and wheat. If BOTANIC or models like it can accelerate trait discovery in these orphan crops, the food security implications are substantial.
Regulatory and practical hurdles remain
Scoring genomic variants computationally is one thing. Getting improved varieties into farmers’ fields is another. Regulatory frameworks for conventionally bred crops are generally less restrictive than those for genetically modified organisms, and genomic selection assisted breeding falls into the conventional category since it guides crossing decisions rather than introducing foreign DNA. This is an important distinction. Models like BOTANIC could accelerate breeding programs without triggering the GMO regulatory apparatus that has slowed adoption of transgenic crops in Europe, parts of Africa, and elsewhere.
However, the gap between a model’s variant score and a validated, field ready crop variety is still measured in years. Genotype to phenotype prediction in plants is complicated by gene by environment interactions, epistasis, and the sheer variability of real world growing conditions. A variant that confers drought tolerance in one soil type and climate zone may behave differently elsewhere. The model can narrow the search, but it cannot eliminate the need for multi environment field trials.
The bigger picture for domain specific AI
BOTANIC fits into a broader pattern that has been emerging across AI development over the past 18 months. The era of building ever larger general purpose models is giving way to a parallel movement toward smaller, domain specific models trained on carefully curated datasets. We have seen this in legal AI, materials science, medical imaging, and now plant genomics. These models often outperform much larger general purpose systems on their target tasks while requiring a fraction of the compute.
The eight GPU training footprint is worth emphasizing. It suggests that the limiting factor for high impact AI applications in specialized domains is not compute but data curation and domain expertise. Living Models reportedly spent considerable effort assembling and quality checking the 1,600 genome training set, which is the kind of painstaking work that does not generate headlines but determines whether a model actually learns meaningful biology or just memorizes sequence statistics.
For investors and strategists watching the AI landscape, this is a signal worth tracking. The next wave of commercially valuable AI may not come from scaling laws applied to internet text but from focused models that speak the language of specific scientific and industrial domains. Agricultural genomics is a trillion dollar global industry that has been slow to adopt deep learning. BOTANIC may not be the model that transforms it overnight, but it establishes a credible proof point that the transformer architecture can meaningfully accelerate the work.
The real test will come over the next two to three years as breeding programs begin incorporating these predictions into their pipelines and we start seeing whether the computational promise translates into varieties that actually perform better under stress in the field. That is the gap between a benchmark and a breakthrough, and it is where the hard work begins.
The window for adapting global agriculture to rising temperatures is narrowing faster than most breeding programs can move. Developing a new crop variety through conventional methods takes anywhere from seven to fifteen years. Climate projections suggest that by the time many of those varieties reach farmers’ fields, the conditions they were bred for may have already shifted. That mismatch between biological timelines and climate velocity is one of the most consequential bottlenecks in food security, and a startup called Living Models thinks it can compress it.
The company has built BOTANIC, a family of transformer models trained not on human language but on the genomic sequences of 43 plant species. The flagship version, BOTANIC 0, was pretrained on more than 1,600 curated plant genomes. It treats DNA as text, learning the grammar of gene regulation, conserved functional motifs, and the associations between sequence patterns and observable traits. And it did all of this on just eight NVIDIA H100 GPUs, which in the current landscape of foundation model development amounts to a rounding error.
That last detail deserves more attention than it might initially receive.
Why Efficiency Is the Real Story
When most people hear about a new foundation model, the instinct is to evaluate it by parameter count and training budget. BOTANIC 0 spans 100 million to 1 billion parameters, which places it firmly in the small to mid range by today’s standards. Living Models raised $7 million, a figure that would not cover a single month of compute costs at the labs training frontier language models. Yet across 22 benchmark tasks in genomics, the model reportedly matches state of the art performance.
This is significant because it challenges a persistent assumption in the broader AI industry: that meaningful progress requires massive capital expenditure on compute. What BOTANIC suggests is that when you have the right domain specificity, careful data curation, and efficient architecture design, you can build models that punch well above their weight class.
The genomics space is narrower than general language modeling, which means the entropy of the data is lower and the signal to noise ratio is higher. A well designed model does not need to learn the entirety of human knowledge to become useful. It just needs to learn the right biology.
This pattern has shown up before. In protein structure prediction, DeepMind’s AlphaFold achieved transformative results not by building the largest model in existence but by combining domain expertise with clever architectural choices. BOTANIC appears to follow a similar philosophy, applied to a different layer of biological complexity.
From Sequence to Stress Tolerance
The core proposition is straightforward in concept if not in execution. BOTANIC functions as a hypothesis engine. Feed it genomic data, and it scores variants for their likely contribution to traits that matter for climate adaptation: drought tolerance, heat resistance, salinity adaptation, nutrient use efficiency. This approach aligns with the transformation of drug discovery by employing AI to streamline the identification of beneficial traits.
The model’s cross species training is particularly important here, because many of the genetic mechanisms that confer stress tolerance exist in wild relatives of cultivated crops but have never been systematically characterized.
Traditional marker assisted selection relies on known associations between specific genetic markers and desirable traits. That approach works, but it is inherently limited to what has already been discovered and validated through years of field research. A language model trained across dozens of species can identify regulatory elements and sequence patterns that classical methods miss, precisely because it is learning representations of genomic function rather than memorizing a catalog of known markers.
This matters for practical reasons. When a breeder wants to develop a wheat variety that can tolerate the combination of heat stress and low water availability expected in South Asia by 2040, they are searching for needle in a haystack combinations of alleles. BOTANIC narrows that search space by predicting which genomic variants are most likely to produce the desired phenotype, allowing researchers to prioritize candidates for experimental validation and field trials.
The Business Model and Competitive Landscape
Living Models has released BOTANIC 0 with open weights, which is a deliberate strategic choice worth examining. Open weights lower the barrier to adoption, allowing seed companies and research institutions to fine tune the model on their own proprietary germplasm data within secure environments.
The value capture likely comes from enterprise licensing, custom fine tuning services, and integration into breeding pipelines rather than from restricting access to the base model. This approach mirrors what Meta has done with Llama in the general purpose AI space and what several biotech startups have attempted in protein engineering. Open weights build ecosystem adoption. Ecosystem adoption creates switching costs. Switching costs create durable commercial relationships even when the foundation is technically open.
The competitive landscape for AI in agricultural genomics is still forming. Companies like Benson Hill have applied machine learning to crop improvement, and major agricultural conglomerates including Bayer, Corteva, and Syngenta have internal computational biology teams with significant resources.
What Living Models offers that those larger players may lack is a purpose built genomic language model with broad phylogenetic coverage, trained specifically to generalize across species rather than optimize for a single crop within a proprietary breeding program.
The question is whether $7 million in funding provides enough runway to establish the platform before a larger player either builds something equivalent or acquires the capability. In agricultural biotechnology, distribution and regulatory relationships often matter as much as technical superiority. A model that can accurately predict climate resilience traits is only valuable if it connects to the organizations that actually develop, test, register, and sell seed.
What People Are Overlooking
Three things stand out that most coverage of this kind of development tends to skip.
First, the regulatory implications are nontrivial. Genome editing technologies like CRISPR are already subject to varying regulatory frameworks across jurisdictions. If AI models begin directing which edits to make, regulators will eventually need to evaluate whether AI guided genomic modifications require different oversight than those identified through conventional research.
The EU, which has recently revised its stance on certain genomic techniques in agriculture, has not yet addressed the role of AI in directing those techniques. This is a gap that will need to close as the technology matures.
Second, the potential for concentrating genetic improvement capability in a small number of AI platforms raises questions about equity in global agriculture. Smallholder farmers in sub Saharan Africa and South Asia, who are most vulnerable to climate impacts, are also least likely to have access to AI guided breeding pipelines. Open weights help, but access to compute, expertise in fine tuning, and the germplasm data needed to make the model useful for locally adapted varieties remain significant barriers.
Third, there is an underappreciated risk in over relying on sequence based trait prediction without sufficient field validation. Genotype to phenotype relationships in plants are notoriously context dependent. A variant that confers drought tolerance in one soil type or climate zone may behave differently in another. Recent independent research underscores this complexity, showing that a single base change in Arabidopsis can affect flowering time by altering the binding of multiple transcription factors simultaneously.
AI can accelerate hypothesis generation, but it cannot replace the multi environment field trials that ultimately determine whether a variety performs under real world conditions. The danger is that enthusiasm for computational prediction creates pressure to shortcut the validation steps that breeding programs exist to perform.
Where This Fits in the Broader AI Trajectory
BOTANIC represents a specific instance of a much larger trend: the application of transformer architectures and language modeling techniques to non language domains. We have seen this in protein structure prediction, molecular design, weather forecasting, materials science, and now plant genomics.
Each application shares a common thesis: that sequence data, whether amino acids, nucleotides, or atmospheric variables, contains learnable structure that transformers are uniquely suited to extract.
What makes the agricultural application distinctive is the urgency of the problem it addresses and the scale of the population it could affect. Climate adaptation in agriculture is not an optimization problem for a niche market. It is a civilizational challenge affecting billions of people.
If models like BOTANIC can meaningfully compress the timeline for developing resilient crop varieties, the downstream impact on food security, land use, water consumption, and rural economies could be substantial.
The next twelve to eighteen months will reveal whether the model’s benchmark performance translates into real predictive value for breeders working on specific crops in specific environments. That translation from benchmark to field is where most AI applications in biology encounter their hardest test.
Living Models appears to understand this, positioning BOTANIC not as a replacement for experimental biology but as a tool for making experimental biology more efficient. Whether that positioning holds as the company scales will depend on partnerships, validation data, and the willingness of an inherently conservative agricultural industry to trust a seven million dollar startup with decisions that affect food production.
The technical foundation looks promising. The execution challenge is everything that comes after.
