Skip to content

OLMo 2

Fully open language models in 7B and 13B sizes

Visit official site

What OLMo 2 is

The model family is described as a fully open release from Ai2 with 7B and 13B sizes trained on up to 5T tokens.

What you can do with it

  • Download model weights from the Hugging Face page
  • Use the Instruct variant to build assistants
  • Evaluate strengths and weaknesses with the OLMES framework

Model notes

  • Staged training uses learning rate annealing and data curriculum changes
  • Architecture changes include RMSNorm, QK-Norm and rotary position embeddings
  • Two-stage pre-training uses OLMo-Mix-1124 and Dolmino-Mix-1124 datasets

Pros & cons

✓ What we like

  • Weights and training data are described as fully open
  • RMSNorm and QK-Norm support stable long pre-training

! What to watch out for

  • English academic benchmarks are used for comparison
  • Large-scale training needs enough compute resources

FAQ

What sizes are offered?

The material lists 7B and 13B sizes.

What is OLMES for?

The material says it sets performance targets and task scaling rules.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review