OLMo 3

weights

open(Apache-2.0)

data

open(Dolma 3, 9.3T tokens, all stages)

code

open(training code+recipes+logs)

checkpoints

open(intermediate)

license

Apache-2.0(OSI)

Fully open: Apache-2.0 weights + the complete Dolma 3 corpus + training code + intermediate checkpoints. The reference fully-open model line.

• https://huggingface.co/api/datasets/allenai/dolma recorded 2026-07-30

  `gated: false`, `private: false`, 2,767 downloads, 1,064 likes, card license `odc-by`. The Dolma corpus is published and ungated.

• https://allenai.org/blog/olmo3 recorded 2026-07-30

  "Olmo 3 is pretrained on Dolma 3, a new ~9.3-trillion-token corpus"; "Because the Olmo 3 data recipes, training pipeline, and checkpoints are open, independent teams can connect model behavior back to measurable" training decisions; "Olmo 3-Base delivers the strongest performance among fully open base models - where training data, code, and weights are all publicly available".

• https://huggingface.co/allenai/Olmo-3-1025-7B recorded 2026-07-30

  OLMo 3 7B base model card: License apache-2.0, safetensors weight files, `isGated: false`.

Succeeds OLMo 2 (the OLMo family measured ~14.8M cumulative HF downloads in early 2026). The reference fully-open model in research/interpretability circles; below the Qwen/Llama scale.

• https://allenai.org/blog/olmo3 recorded 2026-06-30

  flagship fully-open release; broad research uptake

Strongest fully-open base/reasoning model to date; competitive with Qwen2.5/Gemma3 and approaching Qwen3-32B, but still a step below the trillion-param open-weight frontier (=5).

• https://allenai.org/blog/olmo3 recorded 2026-06-30

  OLMo 3-Think 32B strongest fully-open reasoning model; OLMo 3-Base beats fully-open peers (K2, Marin, Apertus)