Tahoe Bio, a U.S.-based life sciences company formerly known as Vevo Therapeutics, has officially unveiled its groundbreaking AI foundation model, Tahoe-x1 (Tx1), a 3-billion-parameter large model specifically designed to decode the complex relationships between genes, cells, and drugs. This release marks a pivotal shift in artificial intelligence—from a supportive tool to a full-fledged modeling engine for biological systems—opening new frontiers in cancer target discovery and personalized therapy development. Built on a Transformer encoder architecture and pre-trained using masked language modeling (MLM), Tahoe-x1 was trained on an unprecedented dataset of 266 million single-cell transcriptomes. This includes Tahoe Bio’s proprietary Tahoe-100M perturbation dataset, which captures the molecular responses of thousands of cancer cell lines to various genetic and pharmacological perturbations—now downloaded nearly 200,000 times by researchers worldwide. To balance performance and accessibility, the model family offers multiple sizes (such as Tx1-70M), and leverages optimizations like FlashAttention v2, enabling computational efficiency 3 to 30 times higher than comparable single-cell models. This allows the model to run efficiently even on standard GPUs, dramatically lowering the barrier to entry for academic and clinical labs. One of Tahoe-x1’s most significant capabilities lies in its precision for identifying cancer vulnerabilities. In benchmark tests on the authoritative DepMap dataset, the model outperformed all existing approaches in predicting gene essentiality—accurately pinpointing the core driver genes that sustain tumor survival across diverse cancer subtypes. This enables researchers to rapidly identify high-value therapeutic targets, drastically shortening the timeline from discovery to validation, particularly for highly heterogeneous and treatment-resistant cancers. Beyond single genes, Tahoe-x1 excels at reconstructing the dynamic molecular networks underlying cancer progression. In tests using the MSigDB database, the model achieved the highest accuracy in identifying “hallmark programs” such as uncontrolled cell cycling and defective DNA repair. By uncovering coordinated signaling pathways, it provides systems-level insights crucial for designing effective multi-target combination therapies. Perhaps the most transformative feature is Tahoe-x1’s zero-shot generalization capability. The model can predict drug responses in cell types or patient samples it has never encountered before, leveraging analogical reasoning based on its vast training knowledge. This enables virtual clinical trials at scale—simulating thousands of treatment combinations in silico before moving to lab or clinical testing. When combined with post-training adaptation frameworks, the model can be fine-tuned for diverse patient populations, accelerating the path to personalized cancer treatments. Tahoe Bio has raised $42 million to date and is building what it calls the world’s largest single-cell perturbation atlas, targeting 1 billion data points. The company is fully embracing open science: model weights, code, and interactive demos are available on Hugging Face and GitHub, while a preprint has been published on bioRxiv. This open, data-driven approach is accelerating the maturation of the biological AI ecosystem. AIbase observes that Tahoe-x1’s true breakthrough lies in its ability to move beyond statistical correlation toward genuine mechanistic understanding. When AI can reason like a biologist—predicting how genes regulate, how drugs intervene, and how cells respond—the paradigm of drug discovery is shifting from trial-and-error to prediction-driven innovation. As data scales continue to grow, Tahoe-x1 has the potential to become a foundational infrastructure for precision medicine: simulating millions of treatment scenarios in the virtual world to ensure the best possible outcome in the real one.