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)

OLMo publishes Apache-2.0 weights, the full Dolma 3 training corpus, training code and intermediate checkpoints, all ungated and downloadable, so the whole training run is open to inspection.

• https://allenai.org/blog/olmo3 recorded 2026-09-29

  Ai2 release post: Olmo 3 is pretrained on Dolma 3, "a new ~9.3-trillion-token corpus drawn from web pages, science PDFs"; "Because the Olmo 3 data recipes, training pipeline, and checkpoints are open, independent teams can connect model behavior back to measurable properties"; fully open means "where training data, code, and weights are all publicly available".

• https://api.github.com/repos/allenai/OLMo recorded 2026-08-13

  `private: false`, `archived: false`, license `Apache-2.0`, 6,625 stars, description "Modeling, training, eval, and inference code for OLMo" - the training pipeline itself, not an inference wrapper.

• https://huggingface.co/allenai/Olmo-3-1025-7B recorded 2026-09-29

  OLMo 3 7B model page states "The code and model are released under Apache 2.0" and serves the Apache-2.0 7B base weights; the embedded repo data records gated as false. Searched for agree to share and access request - none.

• https://huggingface.co/allenai/Olmo-3-1025-7B/raw/main/README.md recorded 2026-09-29

  OLMo 3 7B model card: "We are releasing all code, checkpoints, and associated training details"; "For pretraining, the naming convention is"; lists "Core repo (training, inference, fine-tuning etc.): https://github.com/allenai/OLMo-core" and the pretraining launch script "OLMo-3-1025-7B-pretrain-1.py"; "The code and model are released under Apache 2.0."

• https://huggingface.co/api/datasets/allenai/dolma recorded 2026-09-29

  Hugging Face API record for the original Dolma corpus - "Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research", tagged "license:odc-by"; gated and private are both false.

• https://huggingface.co/api/models/allenai/Olmo-3-1025-7B recorded 2026-09-29

  Hugging Face API record for allenai/Olmo-3-1025-7B tagged "license:apache-2.0" and "dataset:allenai/dolma3_mix-6T-1025", with safetensors shards such as "model-00001-of-00003.safetensors"; gated and private are false.

• https://huggingface.co/api/models/allenai/Olmo-3-1025-7B/refs recorded 2026-09-29

  Hugging Face refs listing for allenai/Olmo-3-1025-7B carries 1,487 branches, including intermediate revisions such as "stage2-step47684-mix-round5-from-2T-ckpt" and "stage3-step10000".

• https://huggingface.co/api/models/allenai/Olmo-3-1125-32B recorded 2026-09-29

  Hugging Face API record for allenai/Olmo-3-1125-32B tagged "license:apache-2.0" and "dataset:allenai/dolma3_mix-5.5T-1125"; gated and private are false.

• https://huggingface.co/datasets/allenai/dolma3_mix-6T/raw/main/README.md recorded 2026-09-29

  Dolma 3 Mix (6T) dataset card: "is the collection of data used during the pretraining stage to train the Olmo-3-1125-32B model"; "Dolma 3 mix is licensed under the Open Data Commons Attribution License v1.0 (ODC-By)."

Adoption is measured as Hugging Face downloads summed across the two declared checkpoints, the 32B and 7B OLMo 3 releases.

• https://huggingface.co/api/models/allenai/Olmo-3-1025-7B recorded 2026-08-13

  102,474 downloads in the trailing 30 days for allenai/Olmo-3-1025-7B

• https://huggingface.co/api/models/allenai/Olmo-3-1125-32B recorded 2026-08-13

  43,671 downloads in the trailing 30 days for allenai/Olmo-3-1125-32B

OLMo 3.1 Think 32B beats Qwen3-32B on AIME 2025 but trails it on GPQA and MMLU, and the line as a whole is competitive with Qwen2.5 and Gemma 3. It remains a step below the trillion-parameter open-weight models at the frontier.

• https://allenai.org/blog/olmo3 recorded 2026-09-29

  Ai2 release post: Olmo 3-Think 32B is described as "narrowing the gap to the best open-weight models of similar scale" such as Qwen 3 32B "while training on roughly 6x fewer tokens"; the 12/12 update says Olmo 3.1 32B Think "brings substantial gains across math, reasoning, and instruction-following benchmarks".

• https://huggingface.co/allenai/Olmo-3.1-32B-Think/raw/main/README.md recorded 2026-09-29

  Olmo 3.1 32B Think model card evaluation table: "| AIME 2025 | 78.1 | 66.2 | 70.7 | 72.5 | 70.9 |" (Olmo 3.1 32B Think 78.1 vs Qwen 3 32B 70.9), "| GPQA | 57.5 | 55.7 | 57.6 | 58.1 | 67.3 |" and "| MMLU | 86.4 | 85.3 | 85.2 | 85.4 | 88.8 |" (Qwen 3 32B ahead on both).