Ai2 released OLMo with weights, training code, and the Dolma corpus. Researchers could inspect data and intermediate checkpoints while studying how training changed the model.
Olmo (Open Language Model)
Also known as: OLMo · OLMoE · Open Language Model
Plain English
Olmo is a family of language models from the Allen Institute for AI. Ai2 publishes training data, code, and intermediate checkpoints so researchers can study how the models were built.
Technical Overview
Olmo releases expose artifacts across pretraining, midtraining, and post-training, with coverage varying by release. Intermediate Checkpoints and data mixtures make comparative experiments possible across training stages. The family includes dense Transformers, the sparse OLMoE branch, and a hybrid branch that combines attention and linear recurrence.
Timeline (4)
2024
Release- Release
OLMo 2 shipped 7B and 13B base and instruction-tuned models. Ai2 published separate recipes and checkpoints for training and adaptation, allowing researchers to compare the effects of each stage.
2025
ReleaseOlmo 3 published base, instruction, and reasoning checkpoints across the 7B and 32B family. The accompanying model flow tied released artifacts to successive stages of development.
2026
ReleaseAi2 released Olmo Hybrid 7B, replacing sliding-window attention layers from Olmo 3 with Gated DeltaNet recurrent layers. The team compared the hybrid and attention-only designs at multiple training scales.
See also (6)
Open full graphOlmo (Open Language Model)