Paris, France and Berkeley, California, USA
March 11, 2026
The startup is building foundation models trained not on text or code, but on DNA - betting that the same deep learning revolution that transformed language and vision is about to do the same for biology. And with genomic data doubling every seven months, they may be sitting on the AI opportunity no one's talking about.
Living Models, a Paris-Berkeley AI company, today announced $7 million in seed funding to develop foundation models for biology. Not chatbots, not code assistants, but AI trained on DNA, RNA, and multi-omics data to understand the fundamental language of living organisms.
Its first model family, BOTANIC, is dedicated to plant biology, and already matches leading work from Cornell and InstaDeep (backed by BioNTech, $100M+ raised) on key benchmarks.
The platform
Living Models trains transformer neural networks on biological sequences rather than human text. The core architecture learns representations of genomic data that can be fine-tuned for specific downstream applications - in the same way that a single language model can be adapted to translation, summarization, or code generation. The goal is a general-purpose biological AI platform: models that understand the language of life across organisms and domains.
Plant biology is the first vertical. The underlying architecture is designed to generalize across biological domains beyond plants, and the company intends to follow the data wherever sequence-driven discovery is bottlenecked by slow, empirical methods.
"Foundation models proved their value in natural language processing by learning deep patterns from large text corpora. We are applying the same transformer architecture to biological sequences (DNA, RNA, multi-omics measurements). Agriculture is our first vertical because the data is abundant and the commercial need is acute, but the models we are building are designed to transfer across biology." Cyril Véran, CEO and co-founder
Why agriculture first
The global seed industry is dominated by a handful of incumbents - Bayer CropScience, Corteva, Syngenta, BASF, and Limagrain - which collectively spend approximately $8 billion per year on breeding research using methods largely unchanged since the 1960s. Development timelines average eight years from initial crosses to commercial variety release. Pathogen resistance genes are typically overcome within 3–5 years of deployment. Current yield improvement rates of roughly 1% annually are insufficient to meet projected 2050 food demand. The 2024 European heatwave alone reduced French wheat yields by an estimated 15%, with agricultural losses exceeding €2 billion.
Plant biology combines three properties that make it an ideal first domain for biological foundation models: genomic data is abundant and largely unrestricted, the commercial need is acute and quantifiable, and the feedback loop between computational prediction and real-world validation is well established through existing breeding infrastructure.
BOTANIC targets the $60 billion global seed market directly. Trained across 43 plant species representing over 60 billion base pairs, it performs trait predictions computationally; reducing the cycle from initial cross to candidate variety from eight-plus years to two to three, while maintaining field validation for phenotypic confirmation.
"Biology is an information problem at every scale, from a single cell to an entire ecosystem. The genomic data exists across many domains; what's been missing is a model architecture capable of learning from it at scale. We start with plants because the data is rich and the breeding cycle is a clear bottleneck, but the same approach applies wherever sequence data meets slow, empirical discovery." Leonard Strouk, CTO and co-founder
Why plants, not humans
Living Models is a biology foundation model company long-term. They're starting with plants because:
- Data access: No HIPAA restrictions. Public + proprietary breeding data.
- Regulation: In pharma, regulation starts before you prove anything works. In agriculture, regulation starts after efficacy is proven. Nobody cares if you kill plants in a greenhouse.
- Validation speed: Test a genetic prediction in 3-6 months (greenhouse) vs. 10+ years (drug trials)
- Climate urgency: Agriculture is getting hammered by climate change today. Not in 2040.
"Every living thing on Earth runs on the same programming language: DNA codes for RNA codes for proteins codes for phenotype," said Bertrand Gakière, VP Biology. "We're not building another chatbot. We're building a model that can read and interpret that code, which is infinitely more useful than predicting the next word in a sentence."
Technical results
Living Models published a technical report demonstrating that BOTANIC (up to 1 billion parameters) achieves competitive performance with state-of-the-art models across 22 standard benchmark tasks — trained on just 8 NVIDIA H100 GPUs. The new capital secures a dedicated 120-GPU NVIDIA B200 cluster, an order-of-magnitude increase in compute that the company expects to translate directly into larger models, higher predictive accuracy, and expansion into genomic domains beyond plants.
Funding syndicate and European presence
The $7 million seed round was led by investors spanning three continents: UC Berkeley's institutional investment arm, Juniper (Silicon Valley), Asterion (Europe), Artesian (Australia), Galion.exe, Pascual, Kima Ventures, and Station F.
The founding team bridges both geographies and disciplines: Véran (UC Berkeley, repeat founder), Strouk (École Normale Supérieure biochemistry; research at UC Berkeley and NYU; previously founded a generative AI company), and Bertrand Gakière (VP Biology, Université Paris-Saclay). The technical team includes PhDs from Huawei Noah's Ark Lab, Owkin (the Paris-based AI drug discovery company, $300M+ raised), Datadog, Mila, and École Normale Supérieure.
Paris operations position Living Models within French governmental priorities for agricultural technology and climate adaptation.
About Living Models
Living Models develops foundation models for biological sequences—DNA, RNA, multi-omics data—rather than natural language. Based in Paris, France and Berkeley, California. Technical report and model weights available at livingmodels.ai.
Research Paper available on BioRxiv: https://www.biorxiv.org/content/10.64898/2026.02.23.706817v1
BOTANIC models available on Hugging Face: https://huggingface.co/living-models
Read our manifesto: The Genome Is the Largest Unread Library in the Universe. We're Building the Reader
Topics
Species
