Omnii

Omnii is a next-generation genomic language model previewed by Radical Numerics, a San Francisco AI lab that launched publicly on June 15, 2026 with a $50 million seed round led by Emergence Capital. Designed to "read, write, and design across all of biology," Omnii learns directly from genomic sequence and is positioned as the successor to Evo 2 — substantially outperforming both the 7B and 40B Evo 2 checkpoints on variant effect prediction without any additional probe training on frozen embeddings.

The model is the first public output of a team with deep lineage in genomic sequence modeling. Radical Numerics was founded by Eric Nguyen (CEO), Michael Poli (Chief AI Scientist), Stefano Massaroli (President), and Armin Thomas (CTO), several of whom developed the Evo and Evo 2 genomic foundation models and the Hyena / StripedHyena long-context architectures. Omnii is explicitly framed as building on the HyenaDNA, Evo, and Evo 2 lineage while unifying three capabilities that were historically pursued separately: scaling, alignment to functional readouts, and multimodal fusion of sequence with genomic annotation tracks.

As of the June 2026 announcement, Omnii is a research preview. No model weights, training code, paper, or public API have been released; Radical Numerics is offering early access to design partners and collaborators, alongside an interactive "Alzheimer's workbench" demonstrating the model's predictions.

#Key Features

• 2 Mbp context window: Omnii processes up to two million base pairs in a single pass, enabling reasoning over large regulatory landscapes and multi-gene loci at single-nucleotide resolution.
• Multimodal sequence + annotation fusion: Beyond raw DNA, Omnii ingests genomic annotation tracks — currently conservation signals from multiple sequence alignment and phylogenetic scores — with native support for variable input compositions at inference time.
• State-of-the-art variant effect prediction: On ClinVar noncoding SNVs, Omnii reaches 0.975 AUROC and 0.735 AUPRC, far above CADD v1.7 (0.909 AUROC, 0.289 AUPRC), and leads on structural variants and complex-trait benchmarks.
• Zero-shot transfer to experiments: Without task-specific training, Omnii recovers experimentally validated functional variants — at nine microglia loci linked to Alzheimer's disease, it ranks the MPRA-validated variant rs2526377 at TSPOAP1 first, where CADD ranks it sixteenth.
• Biosecurity capability: Omnii is reported to achieve state-of-the-art performance at detecting AI-generated or AI-manipulated pathogens, a capability targeted at pandemic surveillance.

#Technical Details

Omnii uses a multi-hybrid backbone combining block convolutions with a dynamic sparse attention mechanism, an architecture that extends the convolution-plus-state-space design lineage of Hyena, StripedHyena, and Evo to a two-million-base-pair context. The model fuses a DNA sequence stream with genomic annotation tracks and supports variable input compositions at inference. As of the preview, Radical Numerics has not disclosed the parameter count, training corpus size, or full pretraining details. Reported benchmark results include 0.975 AUROC / 0.735 AUPRC on ClinVar noncoding single-nucleotide variants, 0.773 AUROC / 0.763 AUPRC on copy-number structural variants spanning 50 bp to 100 kb, and 0.410 AUPRC on TraitGym complex traits (versus 0.284 for CADD). The company also reports a proof-of-concept RNA aptamer design result using HIV reverse transcriptase SELEX data.

#Applications

Omnii targets clinical variant interpretation, where its zero-shot scoring of substitutions, indels, and structural variants can prioritize variants of uncertain significance and surface causal regulatory variants for complex traits and diseases such as Alzheimer's. Radical Numerics is partnering with a diagnostics company to apply the model to early cancer detection (pancreatic and multi-cancer) and with a US national laboratory to pilot Omnii for detecting emerging pathogens, both natural and AI-generated. Additional stated directions include drug target identification and nucleic acid design, illustrated by an early RNA aptamer design demonstration.

#Impact

Omnii arrives as the debut of a well-funded lab built by the original creators of generative genomics, and its headline claim — outperforming the open Evo 2 40B model on variant effect prediction — signals continued rapid progress in genomic foundation models. Its strong zero-shot results on clinically relevant benchmarks and its concrete recovery of a known Alzheimer's functional variant suggest practical value for variant prioritization. The most important caveat is openness: unlike the fully open Evo 2, Omnii is currently a closed research preview with no released weights, code, paper, or API, so its claims await independent reproduction. Its eventual release model and the disclosure of training and architecture details will determine how broadly the research community can adopt and validate it.