Radical Numerics Raises $50 Million Seed to Automate AI Model Design

Radical Numerics Raises $50 Million Seed to Automate AI Model Design

Radical Numerics, a San Francisco AI research lab founded by veterans of Stanford and the Liquid AI founding team, has closed a $50 million seed round led by Emergence Capital — a rare nine-figure-adjacent bet on a company that is trying to automate the design of AI models themselves, and then point that machinery at the most complicated system humans know: biology.

The round, which surfaced in the July 23, 2026 venture roundup tracking recent Emergence, Sequoia and Andreessen Horowitz activity, was joined by Obvious Ventures, Triatomic Capital, Factory and First Spark Ventures. Stripe co-founder Patrick Collison participated as a pre-seed investor. The company says the capital will fund scaling its next generation of models and hiring frontier AI research talent.

The method: models that redesign models

Radical Numerics is easy to mistake for a pure biotech play, but its core engineering thesis is about AI research itself. The lab's stated research direction is "model conversion" — the idea, as the company puts it in its own technical writing, that "we can optimize models at the level of architecture and training objectives without rebuilding entire systems from scratch." Instead of training every new model from zero, Radical Numerics rewires and retrains existing ones, letting it iterate through far more architectural recipes at a fraction of the cost.

The clearest public proof of that approach is RND1, released in October 2025 and billed as the largest open-source diffusion language model to date. RND1 is a 30-billion-parameter sparse mixture-of-experts model with 3 billion active parameters, converted from Alibaba's autoregressive Qwen3-30BA3B checkpoint and continually pretrained for 500 billion tokens until it behaved as a full diffusion model. The company open-sourced the weights, training recipe and inference code, and reported state-of-the-art results among open diffusion models on reasoning, STEM and coding benchmarks. The company has also described a framework, RISA-1, for recursive self-improvement — an engineered loop combining automated agents, human oversight and full observability to accelerate its own research.

That "AI helping design the next AI" pitch is what makes the headline framing fair: the lab's edge is an automated model-design engine. But its founders are explicit that the engine is a means, not the end.

The mission: general biological intelligence

Radical Numerics was founded in 2026 by Eric Nguyen (CEO, a Stanford PhD in bioengineering and AI), Michael Poli (Chief AI Scientist, Stanford PhD and a Liquid AI founding team member), Stefano Massaroli (President, a former Yoshua Bengio postdoc and Liquid AI founder) and Armin Thomas (CTO, a Stanford postdoc under Chris Ré). The same team helped create the field of generative genomics with Evo and Evo 2, the first AI models capable of reading and writing DNA at scale, which landed on the covers of Science and Nature and were later used by outside scientists to generate the first complete AI-designed genome — a bacteriophage harmless to humans.

The company frames its goal as "general biological intelligence": multimodal models that learn directly across DNA, RNA, proteins and beyond, reasoning across every layer of biology at once. Alongside the raise, it previewed Omnii, a next-generation genomic language model that it says sets a new state of the art in identifying causal regulatory variants and, without task-specific training, recovers experimentally validated functional variants at loci tied to Alzheimer's disease. The same model, the company claims, leads in detecting AI-generated or AI-manipulated pathogens.

Radical Numerics describes a deliberate dual mandate — advancing biological design for human health while building the biodefenses to protect against its misuse. It says it is working with a cancer diagnostics company on pancreatic and multi-cancer detection, and with a national lab to characterize pathogens, whether natural or engineered.

CEO Eric Nguyen did not shy from the stakes. "Evo showed that AI can generate DNA and whole genomes; the next generation of models will go further with the ability to control function, and eventually, create entirely new forms of life," he said. "The same models that can help cure disease may also lower the barrier to designing harmful biology. These forces are inseparable. Biology will be the most consequential application of AI."

Investors leaned into that risk-forward positioning. "Most labs bolt safety on at the end. Radical Numerics built it into the foundation," said Gordon Ritter, founder and general partner at Emergence Capital. "They've paired frontier-model capability with real biosecurity expertise to open a scientific field that didn't exist before."

Why it matters — and what to watch

Fifty million dollars is an enormous seed by historical standards, but it is increasingly the going rate for pedigreed AI research teams in 2026, where founder track records and compute needs push first rounds into territory once reserved for Series B. The scientific advisory bench — Microsoft chief scientist Eric Horvitz, Stanford's Chris Ré, Harvard geneticist George Church and former U.S. defense official Andrew Weber — signals a company positioning itself at the intersection of frontier capability and national security.

The bet worth tracking is whether the automated model-design engine actually compounds. If model conversion and recursive self-improvement let a small team out-iterate far larger, better-capitalized labs, Radical Numerics becomes a case study in AI-accelerated R&D. The risk is equally clear from the founders' own words: the tools that design cures can design threats. Expect the next signals to come from Omnii's move beyond preview, the shape of its biosecurity partnerships, and whether "safety built into the foundation" survives contact with commercial pressure.