In reality, nine trillion is a number that resists intuition. Researchers trained an AI model on nine trillion building blocks of DNA, collected from over one hundred thousand species. In reality, no human team could read that much genetic text in many lifetimes. The model learned anyway, one statistical pattern at a time. Its name is Evo 2, and it may change how we read ourselves.
In reality, this guide unpacks Evo 2 in plain language, without the hype. You will see why scientists call it a chatbot for the genome. The caveats get their own section, because genetic predictions stay promises until experiments confirm them.
In this guide
- Evo 2: the 30-second answer
- What you’ll learn
- Why biologists call DNA a language
- What makes a foundation model different
- How Evo 2 learned its craft
- Predicting mutations: the BRCA1 benchmark
- Writing DNA: from reader to designer
- From benchmarks to patients
- How a prediction becomes knowledge
- What Evo 2 still cannot do
- Why this matters beyond genetics
- Where to go next in this series
- Frequently asked questions
- The bottom line
Evo 2: the 30-second answer
Evo 2 is a DNA foundation model from the Arc Institute, built with NVIDIA compute and collaborators at Stanford. In short, in plain terms, it reads DNA the way chatbots read text. Give it a sequence and it predicts what the letters might mean. Meanwhile, it flags mutations that may cause disease and proposes new sequences of its own. The short version is that DNA became a readable language overnight. Keep that framing nearby, because the rest of this guide builds on it.
Key takeaways
- Meanwhile, Evo 2 is a DNA foundation model from the Arc Institute, built with NVIDIA compute and Stanford collaborators.
- The team reports training it on nine trillion nucleotides drawn from over one hundred thousand species.
- It reads sequences up to one million nucleotides at a time, seeing distant relationships inside a genome.
- In team shared benchmarks, it flagged likely harmful BRCA1 variants with strong accuracy.
- Every prediction still needs lab validation, and many populations stay underrepresented in the training data.
What you’ll learn
The route through this guide
- In particular, why scientists call DNA a language
- How the model learned its craft
- What Evo 2 predicts about mutations and disease
- Where the science is already reaching patients
- Which limits deserve your caution
- Where to go next inside this series
Why biologists call DNA a language
In particular, your genome is written in a molecular alphabet, and biologists have long treated it as text. The letters spell genes, the genes shape traits, and small spelling changes can rewrite a health story. In fact, that framing used to be a loose metaphor. Machine learning has turned it into an engineering plan.
In fact, a language model learns the patterns of text by reading enormous amounts of it. Chatbots apply that trick to human writing. Also, Evo 2 applies it to the genetic code, learning which letter combinations tend to follow which. The result behaves, in a rough sense, like a chatbot for the genome. Nothing magical sits underneath, just scale plus pattern learning.
Why does the analogy matter? Because it tells you what the model can and cannot do. In fact, a chatbot predicts likely next words, and Evo 2 predicts likely biological effects. Both can be confidently wrong, which is why validation comes up so often below.
There is a catch hiding inside the compliment, though. In fact, Language models inherit the blind spots of everything they read. A genome database is not neutral, and its gaps become blind spots. Keep that thought handy for the limits section later.
What makes a foundation model different
Most AI tools in biology are narrow specialists. One model predicts protein shapes, another ranks drug candidates and a third matches patients to trials. A foundation model plays a different role. It learns general patterns from massive data, then adapts to many downstream tasks. Generalists are rarer, because they need far more data and care.
Evo 2 aims to be that general reader for biology’s oldest text. The bet is that one model of the genome can serve many jobs, from flagging mutations to designing sequences. Early results support the bet, while the fine print asks for patience. That balance, promise plus patience, runs through everything below.
How Evo 2 learned its craft
The Arc Institute led the project, with NVIDIA providing compute and collaborators at Stanford rounding out the team. Their choice of fuel matters as much as their hardware. The team reported training the model on nine trillion nucleotides, the building blocks of DNA. Feeding a model that much text is a statement of intent.
Those nucleotides came from over one hundred thousand species spread across the tree of life. Breadth was the point. Patterns that repeat across many species tend to mark something biologically important. Training on so much variety teaches the model which changes look tolerated and which look suspicious. Evolution, in other words, was the first editor.
- Collect DNA sequences from over one hundred thousand species across the tree of life.
- Train on nine trillion nucleotides using NVIDIA compute.
- Extend the reading window to one million nucleotides at a time.
- Benchmark the model on tasks such as BRCA1 variant prediction.
Scale shows up in a second way, too. Evo 2 reads sequences up to one million nucleotides at a time. Imagine holding an entire chapter in view instead of a single sentence. That long view lets the model connect far apart signals inside a genome, the distant relationships shorter tools miss.
Why does species breadth matter to a patient reading this? Because biology reuses its best ideas across the tree of life. A pattern seen in one species can illuminate another. Shared grammar is what makes nine trillion letters teachable at all.
Predicting mutations: the BRCA1 benchmark
A mutation is a change in the letter sequence, and most changes are harmless. Some, however, nudge cells toward disease. Sorting dangerous from benign is a central task of modern genetics, and it is slow, expensive lab work. That is where a fast reader earns its place.
This is where the model earned its headlines. In benchmarks the team shared, the model predicted which BRCA1 gene variants are likely harmful with strong accuracy. BRCA1 matters because certain variants raise breast cancer risk, and families live with that uncertainty. The team framed the benchmark as a step toward understanding that risk.
Strong benchmark numbers are still lab numbers. A prediction about a variant is a hypothesis, not a diagnosis. Genetic counselors and clinicians make the calls that touch real people, and they will for a long time.
Benchmarks also have a shelf life. A result shared by the team invites independent checks, and those take time. Treat early numbers as promising signals rather than settled facts. That habit will serve you across the whole AI in medicine beat.
Writing DNA: from reader to designer
Reading is only half of the title’s promise. Evo 2 can also design new genetic sequences. The team reports that researchers tested those designs by building working genome elements. In plain terms, the model proposed DNA text, and lab teams checked whether it functioned. That is an early but striking result.
Design flips the usual workflow. Instead of hunting through nature’s catalog, scientists could request a sequence with a chosen behavior and study it. The applications reach from basic research to potential medicines, though therapies sit far in the future. Biology is complicated, and designs must still earn their proof in experiments. The lab, in other words, stays in the loop.
How should a curious reader weigh this? Hold two ideas at once. Writing DNA is a genuine capability, demonstrated in controlled tests. Using that capability responsibly is a much larger question, and it belongs to institutions, not models.
From benchmarks to patients
Genome models are already brushing against patient care. In 2025, coverage described popEVE, a tool that predicts how likely each genome variant is to cause rare disease. For families stuck without answers, ranking suspicious mutations first can shorten a very long journey. The tool is young, and its reach is still being studied.
Rare disease care saw another signal in 2026. Researchers reported an AI reasoning model helping doctors reach 18 new rare disease diagnoses in previously unsolved cases. These are early reports, and coverage remains thin. Even so, the direction is hard to ignore. Families waiting on a diagnosis will feel this first, if it holds.
Mutation reading also feeds neighboring advances. The same variant knowledge informs personalized cancer vaccines, which teach the immune system about a tumor’s specific changes. Genome models and cancer models are converging on one idea. Understand the letters, then act on them. Progress in one corner of the cluster tends to lift the others.
How a prediction becomes knowledge
The workflow around these models matters as much as the models themselves. For example, a typical loop looks like this. The model ranks a shortlist of suspicious variants or candidate designs. Lab teams test the ranked list first, because a good guess saves months. Speed matters because the alternative is testing everything by hand.
Then the loop tightens. Confirmed findings flow back into training data, and the next round starts smarter. Rejected predictions matter too, since failures mark exactly where the model misreads biology. Science has run this way for centuries. The model just makes the guessing step faster.
Patients rarely see any of this machinery, and that is by design. The point of the loop is to keep unproven guesses away from people. Models speed the guessing, while evidence keeps the gate. Everyone ends up better protected.
What Evo 2 still cannot do
Here is the honest column. Model predictions need lab validation before anyone acts on them, and that bottleneck is not going away. Benchmarks measure a slice of biology, while real cells keep their complexity. Treat every striking number as a starting point. The gap between benchmark and bedside is where most overclaiming lives.
Bias is the second limit. The training data underrepresents many populations, so predictions may transfer poorly to people whose DNA was barely sampled. That is not a small footnote in genetics. It is the difference between a tool that serves everyone and one that serves a sample.
We keep a dedicated tour of what AI in medicine still cannot do, and the genome corner is no exception. Read it before trusting any headline, including ours. Skepticism is not hostility toward the field. It is how the field earns trust.
Reading and writing DNA, before and after
| Task | Before models like Evo 2 | With Evo 2 style models |
|---|---|---|
| Read a genome | Teams studied short stretches one at a time | One pass covers up to one million nucleotides |
| Judge a variant | Lab experiments for each candidate change | Benchmarks flag likely harmful variants first |
| Connect distant signals | Far apart DNA relationships were easy to miss | Long context links segments across a genome |
| Design sequences | Design was slow, manual and piecemeal | Models propose working genome elements for testing |
| Hunt rare disease causes | Cases lingered unsolved without ranking tools | Tools like popEVE rank suspicious variants |
None of these rows retires a scientist. Each one changes where the hours go, away from routine search and toward judgment. The pattern matches what labs report across the wider AI shift. Machines widen the funnel, and people decide what passes through. The table is a map of intent, not a snapshot of every lab.
Why this matters beyond genetics
Genome readers are building blocks for bigger ambitions. Researchers are assembling AI simulations of living systems that aim to predict how biology responds to drugs. A model that understands DNA text is raw material for that effort. The two ideas feed each other. Biology, in that plan, becomes something machines can draft and humans can edit.
For patients, the promise arrives slowly but genuinely. Faster variant interpretation could shorten diagnostic odysseys for families. Researchers get a colleague that has read more biology than any of them. The field, meanwhile, gets a test of whether language style AI transfers to life itself.
Curiosity should stay wide here. Zoom out and the pattern repeats across the cluster, from molecules to cells to trials. Each layer of biology is getting its own reader. Evo 2 happens to be the genome’s.
Where to go next in this series
This series keeps moving, and the natural next stop covers virtual cells. There, the story zooms out from DNA text to simulated biology. Curious how the same pattern plays out beyond biology? Our look at how AI is changing Google search covers models rewriting another old system.
If you want the full map, our AI in medicine cornerstone gathers every stop in the cluster. Each spoke stands alone, so begin wherever your curiosity pulls hardest. The cluster also covers diagnosis, trials and the honest limits, so wandering is encouraged.
One more reason this model matters. Reading DNA well is the front door to medicine’s future. Better mutation predictions shape diagnosis, treatment choice and drug design at once. Few tools touch that many parts of the pipeline. That reach is why foundation models pulled biology’s attention. So far, the attention looks earned and still growing.
Frequently asked questions
What is Evo 2 in simple terms?
It is a DNA foundation model, meaning a large AI trained to read DNA like text. Training covered nine trillion nucleotides drawn from more than one hundred thousand species. Given a sequence, it predicts which changes matter. In one line, it is pattern recognition for the language of life.
Who built Evo 2?
The Arc Institute led the work, with NVIDIA compute and collaborators at Stanford. Its journal paper, by Brixi and colleagues, lays out the methods. It is a useful example of institutes and companies sharing one project.
Can Evo 2 diagnose breast cancer?
No. In team shared benchmarks, it flagged likely harmful BRCA1 variants with strong accuracy. That is a research signal, not a diagnosis. Genetic decisions belong with clinicians and validated tests. Screening and diagnosis are different jobs, and only one is ready.
Did Evo 2 write an entire genome?
No. The team reports that it designed new genetic sequences, and researchers built working genome elements to test them. That is far from designing a whole organism. Designing pieces is not the same as designing life.
What is popEVE?
It is a related tool that predicts how likely each genome variant is to cause rare disease. Coverage in 2025 framed it as a step toward faster diagnosis. Tools like this turn a list of variants into a ranked queue for clinicians.
Does Evo 2 have blind spots?
Yes. Training data underrepresents many populations, so predictions may transfer poorly. Every important call still needs lab confirmation. Treat it as a powerful assistant, not an authority. Ask who was included in the data, and you will ask better questions.
The bottom line
Evo 2 turns a decades old metaphor into working software. DNA is a language, and the model has read nine trillion letters of it. It flags suspicious mutations, proposes new sequences and hints at faster rare disease answers. All of that arrives with an asterisk, because experiments remain the final judge.
Overall, watch this space with interest and with discipline. The biology is genuine, the caution is genuine, and both will grow together. You now speak enough of the language to follow along. The next chapters of this story will be written in both letters and code.
Sources
- Arc Institute, Evo 2 model overview and resources, arcinstitute.org
- Brixi et al., Genome modelling and design across all domains of life, Nature
- NVIDIA, technical blog on building Evo 2 with the Arc Institute, nvidia.com
- popEVE coverage, variant effect prediction for rare disease, November 2025
- Rare disease AI diagnosis coverage, 18 previously unsolved cases, June 2026
- Evo 2 benchmark results and BRCA1 analysis, materials shared by the research team