Cluster · live
The cluster, right now.
504 GPUs training the next Palmyra models, stretching context windows past a million tokens, and running the evaluations that decide what ships. Jobs rotate every few hours, so come back later and the floor will look different.
Connecting to the cluster…
Connecting to the cluster…
A representative view of the kinds of workloads our cluster runs. Job names and telemetry are illustrative and follow a schedule; they are not streamed from the cluster.
What you’re looking at
Pretraining
The biggest jobs on the floor: the next Palmyra generations, scaling-law probes and data-mix experiments.
Long context
Pushing context windows past a million tokens, and checking the model still finds the one fact that matters.
Post-training
Fine-tuning and reinforcement learning for tool use, reasoning and honesty.
Evaluation
Benchmarks, red-teaming and regression checks on every checkpoint before anything ships.
Exploration
Small, strange bets. Most fail. The ones that work become the next paper.
Data & infra
Tokenising, deduplicating and scoring data, plus health checks on the hardware.