Successfully Fine-Tuning 35B MoE: Insights into Model Routing

VectorInst · x · 2026-07-30

Vector's team shared their successful setup for fine-tuning a 35B multimodal MoE on 4 A100s: keeping MoE layers intact while stacking 3 optimizations—4-bit quantization, Flash Attention 2, and a forward hook reducing logit memory from 23.3GB to 0.3GB.

With the model running, they audited the learned routing: layers 1-28 route by modality (audio, video, text), while layers 29-48 shift to semantic routing. Text tokens were highly dynamic, showing a clear shift from perceptual encoding to higher-level reasoning in deeper layers.

Related event: Vector Institute Shares 35B MoE Fine-Tuning Practices(3 posts)→

Original post →

More from Research

Research channel →