OLMO 3: A DEEP DIVE INTO THE FULLY-OPEN LLM

Summary from the TLDR AI newsletter

AI2 published the most comprehensive starting point for open LLM research: all checkpoints, training data, and code for OLMo 3. The walkthrough covers every stage from data mixing through the full SFT/DPO/RLVR post-training stack, including OlmoRL infrastructure that cuts RL training from 15 days to 6 days. RL with random rewards, which worked on Qwen, fails on OLMo 3.