What Is Omnii-1? The Genome Language Model Behind the First AI-Designed Phages

On October 20, 2026, the AI lab Radical Numerics introduced Omnii-1, which it calls the world's most advanced multimodal genome language model. The headline claim is striking: phages designed by the model were given to a human patient with an antibiotic-resistant infection. This guide explains what Omnii-1 actually is, what a genome language model does, its eight applications, and the caveats the headlines skip.

What is Omnii-1?

Omnii-1 is a genome language model: the same transformer technology behind large language models, but trained on biological data instead of human text. Instead of predicting the next word, it learns the patterns in DNA, RNA, proteins, and related molecular signals.

Radical Numerics says Omnii-1 was pretrained natively across 28 biological modalities — spanning DNA, RNA, proteins, chromatin accessibility, transcription factor binding, histone modifications, DNA methylation, splicing, and gene expression, conditioned on cell type and tissue (Radical Numerics). After training on trillions of tokens of genetic and multiomic data, the model learns one shared representation it can adapt to many different biological problems.

The company was founded by members of the team behind Evo and Evo 2, earlier genome models used to design the first bacteriophage genomes created entirely by AI (Business Wire via FinancialContent).

Why a "generalist" model matters

For years, biological AI has been built one problem at a time: separate models for proteins, gene regulation, genetic disease, and molecular design. Radical Numerics argues this is changing — that a single general-purpose model can now outperform specialized models on the tasks those models were built for, echoing how general LLMs overtook narrow NLP systems.

The genome is the natural anchor for this. It serves as both a common reference for molecular biology and a record shaped by billions of years of evolution, so a model grounded in DNA can tie together many downstream signals.

The eight applications

Radical Numerics released Omnii-1 alongside eight applications, each starting from the same base model and adapted through mid-training or post-training. A few of the reported results:

• Omnii Genome. Sets a new state of the art on regulatory-track and variant-effect benchmarks, ahead of specialized models such as AlphaGenome and Borzoi, per the company.
• Omnii Phage. Directed phage design for antibiotic-resistant infections (see below). On randomly held-out E. coli strains, its top two recommendations included a biologically active phage for 92.1% of eligible strains.
• Omnii Genetics. Ranked a functional variant within its top four picks in 90% of evaluated GWAS credible sets, and proposed mechanisms across 153 Parkinson's disease signals.
• Omnii Diagnostics. Predicted a blood-based cancer biomarker (cell-free 5hmC) it never saw in pretraining, reaching 0.854 Pearson correlation on variant effects in the evaluated cohort.
• Omnii Defense. Raised pathogenicity-related protein recall from 78.6% to 95.8% when added to a conventional screening pipeline.
• Omnii Rare Earths, Omnii Antibody, and Omnii Cancer Vaccine round out the set, covering protein design for metal separation, antibody design, and vaccine target selection.

All figures are Radical Numerics' own, reported in its launch post and not yet independently reproduced.

The phage milestone — and its caveats

The result driving the headlines is the phage case. Phages are viruses that infect bacteria, and phage therapy is a promising option when antibiotics fail — but finding a phage that targets a specific patient's infection is slow and hit-or-miss.

Radical Numerics trained Omnii-1 to design viable phages, prioritize variants, and predict which phages match which bacterial strains. In what it calls the world's first clinical use of AI-designed phages, Omnii-1 designs were administered to a patient with an antibiotic-resistant bacterial infection under an authorized, physician-led compassionate-use pathway (Omnii Phage write-up). In lab testing, a cocktail containing the designed variants eliminated the patient's resistant isolate in vitro before administration.

Here is where to stay grounded:

• It is a single patient. Radical Numerics says it cannot disclose the patient's clinical course or outcome, citing privacy — so there is no reported result to evaluate.
• It is investigational. Compassionate use lets patients receive therapies outside clinical trials when no approved option exists; it is not regulatory approval, and the company notes observations from one patient cannot establish general safety or efficacy.
• Not peer reviewed. The company's report has not been peer reviewed or published in a scientific journal, and its benchmark comparisons against AlphaGenome, Borzoi, and others are self-reported, pending independent replication (NYT via GV Wire).

Reporting by Carl Zimmer for The New York Times noted the company said the trial was authorized by the FDA — underscoring that this is real news, but news about a first step, not a proven treatment.

How a genome language model is used in practice

Omnii-1 can be prompted much like an LLM, but for biology: reconstruct a missing measurement, score how a genetic variant changes function, or generate DNA with a desired profile. Crucially, the same DNA can yield different predictions depending on the supplied cell or tissue context — which is closer to how biology actually behaves.

For phage design specifically, the model edits selected regions of an existing genome (such as the receptor-binding tail fiber) while accounting for the surrounding sequence, so it can change host targeting without breaking the machinery the phage needs to replicate.

Why this matters for AI builders

Omnii-1 is a domain example of a broader pattern: one general model, many specialized workflows, each wired up with its own data and evaluation loop. Whether you work in biology, finance, or software, the hard part is rarely the model — it is orchestrating the research, the sources, and the checks around it.

If you want to read launches like this critically, see our explainers on Periodic Neon, a lab-trained AI for science and TypeSafe AI's Jev model.

Run your own research workflow with an AI team

Breaking down a dense launch like Omnii-1 — gathering primary sources, cross-checking benchmark claims, and separating the result from the hype — is exactly the kind of multi-step research an AI workforce can handle. Eigent is an open-source "Cowork" desktop app where a team of agents reads sources, compares claims, and drafts the summary locally, with you approving each step. Point them at a competitive analysis workflow or download Eigent and build your own research desk.