AnalysisAI ModelsAugust 20, 2026

Ornith 1.5 35B A3B ships with untrained MTP head

A HuggingFace discussion reveals Ornith-1.5-35B-A3B is shipped with a Multi-Token Prediction head that was never trained, only randomly initialized, explaining slow performance with MTP enabled.

1 source

More stories today

Open the live feed
Ornith 1.5 35B A3B ships with untrained MTP head — AIBriefs