OpenModelsHQ

Allen Institute for AI

Seattle, USA

Allen Institute for AI is based in Seattle, USA. We track 16 open-weight models from it, and the newest is Bolmo-1B-Stage1, released 26 Aug 2026.

Models tracked
16
LLM 11 · Vision 4 · Embedding 1
Newest release
Bolmo-1B-Stage1
26 Aug 2026
Country
USA
Releases timeline
16 Allen Institute for AI releases from 16 Feb 2023 to 26 Aug 2026. The largest is Molmo2-8B at 8.7B parameters. Dot area follows total parameters; a ring means the size is not published.
Releases timeline: data
ModelCategoryReleasedParameters
Bolmo-1B-Stage1LLM26 Aug 20261.5B
Molmo2-O-7BVision14 Dec 20257.8B
Molmo2-8BVision14 Dec 20258.7B
Olmo-3-7B-InstructLLM19 Nov 20257.3B
Olmo-3-7B-ThinkLLM18 Nov 20257.3B
Olmo-3-7B-Instruct-SFTLLM17 Nov 20257.3B
olmOCR-2-7B-1025Vision6 Oct 20258.3B
Olmo-3-1025-7BLLM12 Sep 20257.3B
OLMo-2-0425-1B-InstructLLM29 Apr 20251.5B
OLMo-2-0425-1BLLM17 Apr 20251.5B
OLMoE-1B-7B-0125-InstructLLM27 Jan 20256.9B
olmOCR-7B-0225-previewVision15 Jan 20258.3B
OLMo-2-1124-7B-InstructLLM18 Dec 20247.3B
OLMoE-1B-7B-0924LLM20 Jul 20246.9B
wildguardLLM15 Jun 20247.2B
specter2_baseEmbedding16 Feb 2023Not published

Origin

Paul Allen, the Microsoft co-founder, set up the Allen Institute for AI in 2014. Ai2, as it now calls itself, is a 501(c)(3) non-profit research institute based in Seattle, with an office in Tel Aviv. It was doing AI research long before open language models: the Aristo reasoning project, the Semantic Scholar search engine (public since November 2015) and the AllenNLP library all came from there.

Ali Farhadi became CEO in July 2023. In March 2026 Ai2 announced that he would step down, with Peter Clark taking the role on an interim basis.

How it got here

Ai2 announced OLMo on 11 May 2023 and released the 1B and 7B versions in February 2024. The weights came with the training code, intermediate snapshots and logs, and Dolma, a three-trillion-token English training corpus. Releasing data and code as well as weights is what Ai2 calls “fully open”.

OLMoE came in September 2024: a mixture-of-experts model with about 7 billion total and 1 billion active parameters, built with Contextual AI. OLMo 2 arrived in November 2024 at 7B and 13B, and a 32B version in March 2025. Olmo 3, released in November 2025, came in 7B and 32B sizes with Base, Think, Instruct and RL Zero variants. An updated 32B, Olmo 3.1, landed in December 2025.

Not everything is a language model. The olmOCR 2 document reader appeared in October 2025 and the Molmo2 vision-language models in December 2025.

Licensing pattern

Every Ai2 card we checked lists Apache 2.0: OLMo 2 32B Instruct, Olmo 3 7B Think, Molmo2 8B and olmOCR 2. The Olmo 3, Molmo2 and olmOCR 2 cards add that the models are intended for research and educational use, in line with Ai2’s Responsible Use Guidelines.

Molmo2 comes with a catch. Its card says it was trained on third-party datasets that are subject to academic and non-commercial research use only, and tells users to review those sources before relying on the Apache 2.0 label for commercial work.

The licence is only half of Ai2’s openness claim. It pairs the weights with training data, code and intermediate checkpoints, and the OLMo 2 card says checkpoints from its reinforcement learning stage were released to help research on that step. We found no change of licence across these releases.

Release lineage

Newest first
  1. Bolmo-1B-Stage11.5B
  2. Molmo2-O-7B7.8B
  3. Molmo2-8B8.7B
  4. Olmo-3-7B-Instruct7.3B
  5. Olmo-3-7B-Think7.3B
  6. Olmo-3-7B-Instruct-SFT7.3B
  7. olmOCR-2-7B-10258.3B
  8. Olmo-3-1025-7B7.3B
  9. OLMo-2-0425-1B-Instruct1.5B
  10. OLMo-2-0425-1B1.5B
  11. OLMoE-1B-7B-0125-Instruct6.9B
  12. olmOCR-7B-0225-preview8.3B

Models from Allen Institute for AI

Compare in the table

Largest first. The last line on each tile is the memory of the smallest device that fits the model at 4-bit.