Nebius at NDSML Summit 2026

Frontier open-weight models such as Kimi K3 and GLM give any team a strong starting point without training from scratch. Post-training that model on your own data — so it knows your domain and stays under your control — is where most teams stall.

This session covers what that takes at trillion-parameter MoE scale: the GPU footprint before the first useful iteration, how trainer and rollout workers share hardware, weight synchronisation, environments for reinforcement learning, and evaluation that tells you the run worked.

We’ll also cover when supervised fine-tuning or LoRA is enough and when RL is needed. Examples come from post-training work with design partners on Nebius AI Cloud. You leave with a realistic picture of the effort, the hardware, and what you own when the run finishes.

Let's meet in person!