How they use Nebius AI Cloud

Stanford’s Le Cong lab and collaborators used Nebius AI Cloud GPU clusters to develop CRISPR-GPT, an AI assistant for gene-editing experiments. Its architecture combines an LLM planner, task executors, user-proxy agents and external bioinformatics tools. Nebius describes training CRISPR-Llama3 on 11 years of scientific-forum discussions and experimenting on an eight-H100 node. The paper describes curating approximately 4,000 discussion threads into over 3,000 question-and-answer pairs to fine-tune an eight-billion-parameter Llama3-instruct model. Parallel model testing reportedly saved several months of development time. Preconfigured libraries and drivers allowed workload deployment without infrastructure setup.

Nebius’s podcast representative also describes training and running the project’s models on Nebius infrastructure. The team developed Gene-editing bench, a 288-case evaluation framework covering experiment planning, guide RNA design, delivery-method selection and Q&A. Human experts supplied ground-truth annotations for benchmarking against existing tools and general-purpose LLMs. Nebius reports that a novice PhD student achieved approximately 80% editing efficiency across four targets on the first attempt. An undergraduate achieved 56.5% and 90.2% efficiency on two separate targets. Nebius lists the paper among research supported by its grants programme.

Le Cong later described collaboration with Nebius on LabOS’s AI and XR laboratory research and Nebius support for the LabSuperVision benchmark.