What's new in the model catalog: 174 open-weight models
Model training4 mins read
Introduction
The catalog now lists 174 open-weight models, each tuned for training and inference and tested for correctness before it's available.
Highlights
New mixture-of-experts models bring large total parameter counts with modest active parameters, which keeps fine-tuning and serving costs predictable.
Every listing shows parameter counts, and MoE models list both total and active parameters, so comparisons stay honest.
Choosing a base model
Start from the smallest model that clears your evaluation set. Larger models are easier to prompt but harder to serve; a well-tuned smaller model often wins in production.
