Soofi S faces overtraining criticism after topping open benchmarks

Days after a German consortium released the open 30B model Soofi S, critics argued it was heavily overtrained by Chinchilla scaling-law standards, using about 27 trillion tokens for 30 billion parameters, hundreds of times past the sweet spot. Counting only the 3.2 billion parameters active per token pushes the ratio into the thousands. The project's technical lead rejected the criticism, arguing classic dense-model scaling laws do not carry over to mixture-of-experts architectures. The model activates 3.2 of its 31.6 billion parameters per token and outperforms Olmo 3 32B and Apertus 70B on German, English, and programming tasks.

View full digest for July 17, 2026