Hero Run Scheduled: AI Virtual Cell Model Training on 192 GPUs
Begins Tuesday, September 8, 2026, 10:00 AM PDT
Professor Jure Leskovec’s team will hold up to 24 nodes (192 H100 GPUs) for a hero and scaling run from Tuesday, September 8 at 10:00 AM PDT through Tuesday, September 15 at 10:00 AM PDT.
The run pretrains a protein language model, part of the group’s work on the molecular foundation of the AI Virtual Cell: reasoning architectures built for biological data modalities (DNA, RNA and proteins). The group has demonstrated scaling laws starting at 770 million parameters. This run extends that work to models with billions of parameters, with the goal of an open-source release.
As announced to Marlowe users on September 2: for the duration of the run the batch partition is down but continues to accept job submissions to the queue, and jobs running in batch when the run begins are requeued and eligible to resume from 10:00 AM on September 15. The preempt partition continues to run jobs as usual.