Radical Numerics

Radical Numerics Inc. is an American artificial intelligence research company based in San Francisco that trains large models on biological sequence and measurement data. It was founded in 2025 by Eric Nguyen, Michael Poli, Stefano Massaroli and Armin W. Thomas, four researchers from the line of work that produced the Hyena sequence architecture and the Evo family of genome models. ^[1][2][4] The company states its objective as "general biological intelligence," which it defines as AI that can interpret, generate and predict biology, and says it pursues biological design and biological defense together rather than separately. ^[1][7]

The company left stealth on 15 June 2026 with a $50 million seed round led by Emergence Capital, announced alongside a research preview of its genome model, Omnii. ^[2][3][4][5] On 20 October 2026 it released Omnii-1 and published eight application studies built on it. ^[6] The most widely covered of these was the company's report that phage variants designed with the model had been given to a patient with a drug-resistant infection, a claim covered by the New York Times, which noted that the report had not been peer reviewed and that the patient's outcome had not been disclosed. ^[8][9]

Founding and research lineage

The founders' shared history runs through the long-context architecture work begun at Stanford University and through the Evo genome models, which the company describes as a series of unconventional technical bets. ^[1] Poli, Massaroli and Nguyen were among the authors of the February 2023 Hyena paper, which proposed replacing attention with long implicit convolutions and gating. ^[22] Nguyen, Poli, Thomas and Massaroli then co-authored HyenaDNA, a DNA model with context lengths up to one million tokens. ^[20] Nguyen was first author and Poli second author on Evo, published in Science in November 2024, with Thomas also a co-author. ^[19] Poli and Nguyen were among the authors of Evo 2, the larger successor published in Nature in 2026 and developed with the Arc Institute and collaborators. ^[21]

According to Fortune and the launch announcement, three of the four founders previously worked at Liquid AI, the architecture company spun out of MIT, and Poli was earlier a senior member of technical staff at Together AI. ^[1][2][4] Nguyen completed a PhD in bioengineering and AI at Stanford, was advised in part by Christopher Ré and Stephen Baccus, and spent the summer of 2025 as a research intern on the AI co-scientist team at Google DeepMind; his own site dates the company's founding to July 2025 and announced it that August. ^[23] He gave a TED talk on AI and genome design that was recorded in April 2025 and published on TED.com that August. ^[28]

The company's public positioning changed between its founding and its launch. Pages captured by the Internet Archive in October 2025 presented Radical Numerics as a lab building "the engine for recursive self-improvement," under the slogan "AI designs AI," with the stated ambition of automating architecture design, data selection and training pipelines; a January 2026 capture described it as an "AI lab for scientific world models." ^[24] By June 2026 the public material described a biology lab, and the founders called the demonstration that a genome model could generate a working viral genome "a clear turning point"; Fortune, which dated that milestone to September 2025, reported that it was what pushed the team to build a company. ^[2][3] That demonstration came from a separate Stanford and Arc Institute team, which used Evo 1 and Evo 2 to generate bacteriophage genomes and reported 16 viable designs; the work appeared in Science in August 2026. ^[9][27]

Funding

The seed round was led by Emergence Capital, with participation from Obvious Ventures, Triatomic Capital, Factory and First Spark Ventures. ^[2][4] Patrick Collison, the chief executive of Stripe and a co-founder of the Arc Institute, invested at the pre-seed stage. ^[2] Neither the company nor the outlets that covered the round reported a valuation, and the pre-seed amount was not disclosed. ^[2][3][4] Gordon Ritter, founder and general partner at Emergence Capital, said in the launch release that the firm led the round because the team "paired frontier-model capability with real biosecurity expertise." ^[4]

In the launch post the founders said the company had a data center under construction using NVIDIA Blackwell hardware. ^[3] Fortune reported that the revenue model was still taking shape and was expected to combine API licensing, models fine-tuned for pharmaceutical partners, and milestone payments, quoting Nguyen's view that nobody had yet worked out how AI companies commercialize in the life sciences. ^[2]

Architecture and systems research

Alongside its biology work the company has published on model architectures, numerics and pretraining data, some of it before Omnii was announced.

• RND1 (October 2025) is a diffusion language model with 30 billion parameters in a sparse mixture-of-experts layout and about 3 billion active parameters, converted from the autoregressive Qwen3-30B-A3B checkpoint and continually pretrained for 500 billion tokens. The company described it as the largest open-source diffusion language model at the time and released the weights on Hugging Face under the Apache 2.0 license, together with a technical report and code. ^[11][12]
• Phalanx (October 2025) is a layer based on sliding window recurrences, offered as a replacement for sliding window attention, with a custom kernel the company calls block two-pass. It reported the highest throughput in its comparison set across sequence lengths from 32 to 512,000 tokens. ^[13]
• Two posts in January 2026 documented NVIDIA's NVFP4 four-bit pretraining recipe and its implementation in TransformerEngine. ^[25]
• Tessera (October 2026) scales pretraining data mixing from a handful of hand-defined sources to thousands derived from document behavior, applied to both language and DNA pretraining. ^[15]
• Halcyon (October 2026) is a family of early-fusion text-to-image models from 0.1 to 4 billion parameters, trained from scratch on 100 million image and text pairs with no language pretraining; the company published checkpoints on Hugging Face and presented the work as a controlled study of how to combine modalities, motivated by the heterogeneity of biological data. ^[14]

Omnii

Omnii is the company's genome model family. The June 2026 research preview described a multi-hybrid backbone combining block convolutions, a generalization of scalar convolutions in which each block is a learned dense matrix, with a context-dependent sparse attention mechanism and a context window of up to 2 million base pairs. ^[5] The company said Omnii was the first genome model to be mid-trained and post-trained, so that it could be used directly rather than through task-specific classifiers fitted on frozen embeddings, and reported state-of-the-art results for genome models on clinical variant classification and on the TraitGym Complex benchmark, with the largest gains on noncoding variants. ^[5] It also described a chain-of-thought style post-training format over trajectories of increasing measured fitness, used in an RNA aptamer design test, and probes of the model's internal representations. ^[5][26]

Omnii-1, released on 20 October 2026, extends the model to measurement data aligned to the genome. The accompanying press release said the company had developed a way to convert 28 types of biological data, including DNA, RNA, proteins and epigenetic signals, into a common representation. ^[6][10] Parameter counts and training data volumes for Omnii-1 have not been published; the company describes the training set only as trillions of tokens of genetic and multiomic data, and the model is available by request rather than as open weights. ^[6] Fortune reported in June 2026 that future model releases would not automatically be open source, in contrast to the fully open Evo releases. ^[2]

On benchmarks for the regulatory genome, the company reported that Omnii-1 outperformed AlphaGenome, Borzoi and ChromBPNet on most variant-effect and track-reconstruction comparisons it ran, and noted that as a commercial entity it could not run AlphaGenome itself and therefore took AlphaGenome's figures from the published paper. ^[29] All of these results come from the company's own technical posts and have not been independently replicated or peer reviewed.

Applications published with Omnii-1

The phage therapy report

The announcement that drew the most attention was the company's statement that phage variants designed with Omnii-1 had been administered to a patient with an antibiotic-resistant bacterial infection under an authorized, physician-led compassionate-use pathway, which it called the first clinical use of AI-designed phages. ^[8] The company said a cocktail containing the designed variants eliminated the patient's bacterial isolate in laboratory testing before administration, that it would not disclose the patient's clinical course or outcome in order to protect privacy, and that observations from a single patient cannot establish safety or efficacy. ^[8]

The New York Times, whose account was syndicated by other outlets, reported that the designs were variants of T7, a long-studied phage that infects Escherichia coli, that the patient's infection had resisted both antibiotics and the available natural phages, and that company officials said the trial had been authorized by the Food and Drug Administration. ^[9] The report also noted that the company declined to identify the patient or the treating physicians, citing confidentiality, and that the technical write-up had not been peer reviewed or published in a journal. Nguyen said the company intended to publish its findings through peer review, with any disclosure of clinical details subject to consent. ^[9]

Outside researchers quoted in that coverage were cautious. Michael Baym, a microbiologist at Harvard University who was not involved, said that the result would be remarkable if it worked but that he would reserve judgment until the treatment was finished and the novelty of the designs could be assessed, adding "My impulse is to be skeptical." ^[9] Marisa Azad, an infectious disease clinician at The Ottawa Hospital who is working on a collaboration with the company, described the search for matching phages as the time-consuming step that the approach aims to shorten. ^[9][10] Later coverage noted that phage therapy outcomes are typically followed for months, that laboratory activity does not establish clinical benefit, and quoted the evolutionary microbiologist Brandon Berryhill to the effect that laboratory experiments cannot reliably predict how a designed phage will behave during a human infection. ^[9][31]

Biosecurity posture, staff and advisors

The company frames its work around what it calls a dual mandate: that the capabilities which help design medicines also create risks, so the same organization should build defensive capability. ^[1][7] In practice this has meant offering the biodefense model through a restricted-access interface for vetted researchers and agencies rather than as an open release, publishing pathogen-screening results in that restricted form, and recruiting national security staff and advisers. ^[2][6][35]

Christine Parthemore, previously chief executive of the Council on Strategic Risks and a senior adviser at the U.S. Department of Defense, joined as head of biodefense in October 2026. ^[18] Jacob Rinaldi, previously at Genentech and Recursion and a co-founder of Noetik, joined as head of scientific strategy in September 2026. ^[17] The advisory group listed on the company's site comprises George Church of Harvard Medical School, Tri Dao of Together AI and Princeton, retired Major General Paul Friedrichs, Eric Horvitz of Microsoft, Jill Hruby, formerly of the National Nuclear Security Administration and Sandia National Laboratories, Christopher Ré of Stanford, Ben Spector of Flapping Airplanes, and Andrew Weber, a former U.S. assistant secretary of defense for nuclear, chemical and biological defense programs. ^{[1][2][4][16][17]}

The company is headquartered in San Francisco and lists research roles in both San Francisco and Tokyo; in late October 2026 its careers page showed 17 open positions. ^[4][34] The recruiting site Built In lists the company at about 25 employees, a third-party figure the company has not confirmed. ^[32] Published collaborations include ClearNote Health on a cell-free DNA biomarker and Lawrence Livermore National Laboratory on protein design for rare-earth separation. ^[6][30]

Timeline

See also

• List of neolabs
• Neolab
• Evo 2 (genomic model)
• Hyena
• Liquid AI
• Arc Institute
• AI for Science
• AI in drug discovery

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Radical Numerics, "About Radical Numerics," accessed 27 October 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Lily Mae Lazarus, Fortune, "Exclusive: The researchers who built AI-generated DNA just raised $50 million to reinvent biology," 15 June 2026.
3. ^1 ^2 ^3 ^4 ^5Eric Nguyen, Michael Poli, Stefano Massaroli and Armin Thomas, Radical Numerics, "Radical Numerics raises $50M seed to build general biological intelligence," 15 June 2026.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Business Wire, "AI Lab Radical Numerics Launches with $50M Seed Round To Build General Biological Intelligence," 15 June 2026 (as republished by CRISPR Medicine News).
5. ^1 ^2 ^3 ^4 ^5Radical Numerics, "A new frontier in generative genomics with Omnii," 15 June 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Radical Numerics, "Introducing Omnii-1: a leap toward general biological intelligence," 20 October 2026.
7. ^1 ^2Radical Numerics, "The Technology: Next-Generation Multimodal Genome Language Model," accessed 27 October 2026.
8. ^1 ^2 ^3 ^4 ^5Aniketh Iyengar, Benjamin Siranosian, Arpan Sarkar, Avery Noonan, Michael Poli, Eric Nguyen and others, Radical Numerics, "Radical Numerics creates first AI-generated phages to reach human patient with Omnii-1," 20 October 2026.
9. ^1 ^2 ^3 ^4 ^5 ^6 ^7Carl Zimmer, The New York Times, "Startup claims to use AI-created viruses to treat a patient," 20 October 2026 (as republished by The Boston Globe).
10. ^1 ^2Business Wire, "Radical Numerics Launches Omnii-1, a New Generation of Biological AI, as First AI-Designed Phage Therapy Reaches Human Patient," 20 October 2026 (as republished by 01net).
11. ^1 ^2Keshigeyan Chandrasegaran, Armin W. Thomas and others, Radical Numerics, "RND1: simple, scalable AR-to-diffusion conversion," 9 October 2025.
12. ^1 ^2Hugging Face, "radicalnumerics/RND1-Base-0910," model card, created 4 October 2025.
13. ^1 ^2Garyk Brixi, Dragos Secrieru, Stefano Massaroli and others, Radical Numerics, "Phalanx: hardware-aligned sliding-window recurrences," 14 October 2025.
14. ^1 ^2Keshigeyan Chandrasegaran, Yue Zhao, Armin W. Thomas and Michael Poli, Radical Numerics, "Text-to-image generative modeling without language pretraining," 15 October 2026.
15. ^1 ^2Willie Neiswanger, Matthew Carter, Anvith Thudi and Michael Poli, Radical Numerics, "Scaling data mixing for language and DNA pretraining," 12 October 2026.
16. ^1 ^2Eric Nguyen, Radical Numerics, "Tri Dao joins Radical Numerics as scientific advisor," 16 October 2026.
17. ^1 ^2 ^3 ^4Eric Nguyen, Radical Numerics, "Radical Numerics Welcomes Jill Hruby and Paul Friedrichs as Biodefense Advisors," 19 October 2026; and Radical Numerics, "Jacob Rinaldi joins Radical Numerics as Head of Scientific Strategy," 8 September 2026.
18. ^Christine Parthemore, Radical Numerics, "Christine Parthemore on joining Radical Numerics as Head of Biodefense," 5 October 2026.
19. ^1 ^2Eric Nguyen, Michael Poli, Matthew G. Durrant, Armin W. Thomas and others, "Sequence modeling and design from molecular to genome scale with Evo," *Science*, 15 November 2024.
20. ^1 ^2 ^3Eric Nguyen, Michael Poli, Marjan Faizi, Armin Thomas, Stefano Massaroli and others, "HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution," arXiv, 27 June 2023.
21. ^Garyk Brixi, Matthew G. Durrant, Jerome Ku and others, including Michael Poli and Eric Nguyen, "Genome modelling and design across all domains of life with Evo 2," *Nature* 652, 1349-1361, 2026.
22. ^1 ^2 ^3Michael Poli, Stefano Massaroli, Eric Nguyen, Daniel Y. Fu, Tri Dao and others, "Hyena Hierarchy: Towards Larger Convolutional Language Models," arXiv, 21 February 2023.
23. ^1 ^2 ^3Eric Nguyen, "Portfolio: updates, education and work experience," accessed 27 October 2026.
24. ^Internet Archive, "Radical Numerics: about page, capture of 10 October 2025," and home page captures of 13 October 2025 and 15 January 2026.
25. ^1 ^2Jerome Ku, Michael Poli and others, Radical Numerics, "NVFP4 pretraining: from theory to implementation (Part 1)," 12 January 2026.
26. ^Latent Space, "Bio-security is an AI Arms Race: Eric Nguyen (CEO, Radical Numerics)," 23 September 2026.
27. ^Samuel H. King, Claudia L. Driscoll, David B. Li, Daniel Guo, Aditi T. Merchant, Garyk Brixi, Max E. Wilkinson and Brian L. Hie, "Generative design of bacteriophages with genome language models," *Science* 393, 6 August 2026.
28. ^1 ^2TED, "Eric Nguyen: How AI could generate new life forms," recorded April 2025, published 25 August 2025.
29. ^1 ^2Matthew Carter, Boris Meinardus, Willie Neiswanger, Michael Poli and Armin Thomas, Radical Numerics, "Generative multiomic modeling with Omnii-1," 20 October 2026.
30. ^1 ^2Armin Thomas, Gautam Machiraju, Benjamin Siranosian and Christine Parthemore, Radical Numerics, "Finding proteins that separate rare earth metals, with Omnii-1," 20 October 2026.
31. ^Bal Marsius, International Business Times UK, "Could AI Kill Superbugs? Company Says AI-Designed Phages Given to First Patient, but Outcome Remains Unknown," 26 October 2026.
32. ^Built In Los Angeles, "Radical Numerics company profile," accessed 27 October 2026.
33. ^Michael Poli, "Personal homepage," accessed 27 October 2026.
34. ^Radical Numerics, "Join Us," accessed 27 October 2026.
35. ^1 ^2Gautam Machiraju, Benjamin Siranosian, Avery Noonan, Christine Parthemore, Stefano Massaroli and others, Radical Numerics, "Biodefense with Omnii-1, through API and edge deployment," 20 October 2026.