Poolside ships a 118B model with under 70 researchers

Poolside built a 118B-parameter mixture-of-experts model in eight weeks, co-CEO Eiso Kant told Latent Space. The model has 8B active parameters per token and a 1M context window, and Poolside's team is under 70 researchers. That ratio is the story: Kant says the team runs 10,000 to 20,000 experiments a month without headcount to match.

Checkpoints are evaluable within 30 minutes of completion, per Kant. Training data streams in instead of getting pre-materialized to disk, which kills the re-staging cycles that slow most large runs.

The training run itself used a 10,000 H200 cluster and produced zero on-call incidents across the full eight weeks, Kant said. Laguna S 2.1 shipped as open weights under OpenMDW-1.1. It fits on a single DGX Spark with FP8 precision, a real constraint check on a 118B model.

Kant's thesis: model building is 90% engineering, and 95% of that reduces to improving data or compute efficiency. He doesn't say how the 30-minute eval holds up on harder capability benchmarks versus quick sanity checks. He also doesn't say what the zero-incident run cost in infrastructure work before the eight weeks started.