Allen Institute for AI

Ai2 (Allen Institute for AI) is a lab that publishes open AI models, such as the Olmo language models and the Molmo multimodal models, together with training data, code and checkpoints.

According to Ai2's website, the main line is:

• Olmo: language models. Ai2 calls the full lifecycle of a model the "model flow," with every stage, checkpoint, dataset and dependency needed to create and modify it. The Olmo 3 blog describes a family of dense models with 7B and 32B parameters.
• Molmo: a family of open multimodal models, backed by the Pixmo dataset. Molmo 2 includes 4B and 8B models and Molmo 2-O, at 7B, which pairs Molmo 2's image and video understanding with Olmo.
• Tülu: a family of instruction-following models with fully open data, code and recipes. Ai2 applied the Tülu 3 methods to build the Olmo 2 Instruct models.

Ai2 states that it shares the training data, code, methods, checkpoints, evaluations and documentation behind the models. Olmo 3.1 models can be tried in the Ai2 Playground or downloaded from Hugging Face.

Ai2 is for researchers, ML teams and educational institutions that want to study, reproduce or adapt AI models with access to training data, code, checkpoints and documentation.

Use cases that Ai2's website and model pages support:

• Research on how a language model is trained, using Olmo's "model flow," with pretraining, mid-training data and checkpoints. The Olmo 2 paper cites thousands of intermediate checkpoints.
• Adapting models to a task, starting from the Olmo 3 Base, Think (reasoning) and Instruct (chat) versions, or from Tülu's recipes and data.
• Applications with images, video and multiple images, using the Molmo 2 models.
• Quick testing in the Ai2 Playground, without downloading the weights.

What to have before you start:

• Attention to intended use. Olmo's cards on Hugging Face state use for research and education in accordance with Ai2's Responsible Use Guidelines, even with the Apache 2.0 label.
• For Molmo 2, read the card's notice: the models were trained on third-party datasets subject to academic and non-commercial research terms.
• Your own infrastructure to run the downloaded weights.

Ai2's main offerings are the Olmo language models, the Molmo multimodal models and the Tülu post-training recipes, all published with data, code and documentation.

• Olmo 3: a family of dense models with 7B and 32B parameters, according to Ai2's blog. Hugging Face lists Base, Think and Instruct versions, and Olmo 3.1 brings 32B Think and Instruct checkpoints.
• Model flow: Ai2 publishes every stage, checkpoint, dataset and dependency in the model's lifecycle, including pretraining and mid-training mixtures.
• Olmo 2: 7B, 13B and 32B models published with weights, training data, code, recipes, logs and thousands of intermediate checkpoints, according to Ai2.
• Molmo 2: open vision-language models that support images, video and multiple images. The 7B Molmo 2-O is built on the 7B Olmo 3 Instruct.
• Pixmo: the multimodal dataset behind Molmo.
• Tülu: open data, code and recipes for instruction-following models.
• Ai2 Playground: an environment for trying the Olmo 3.1 checkpoints in the browser.

Ai2 also publishes the model documentation at docs.allenai.org.

Ai2 is worth it for anyone who needs to see inside the model: data, code, checkpoints and documentation are published together with the weights. For anyone who needs unrestricted commercial use, read the model card notices first.

Pros:

• Ai2 publishes Olmo's complete "model flow": stages, checkpoints, datasets and dependencies.
• Olmo 2 came out with weights, training data, code, recipes, logs and thousands of intermediate checkpoints, which allows reproducing training stages.
• Olmo 3 has Base, Think and Instruct versions, for fine-tuning, reasoning or chat.
• Molmo 2 accepts images, video and multiple images, and Molmo 2-O is built on an Olmo model.

Cons:

• Olmo's cards on Hugging Face state use for research and education, per Ai2's Responsible Use Guidelines, in addition to the Apache 2.0 label.
• Molmo 2's cards warn that training used third-party datasets subject to academic and non-commercial terms.
• Olmo 3 comes in two sizes, 7B and 32B, according to Ai2's blog.

The main alternatives to Ai2 as an open AI model provider are NVIDIA, IBM Granite, AI21 Labs and Amazon Nova.

• NVIDIA: the open Nemotron model family, with weights on Hugging Face, and NeMo, a suite for curating data, fine-tuning models, retrieving information and applying guardrails. NVIDIA also publishes articles about Nemotron's open models, data and recipes.
• IBM Granite: language, vision, speech, embedding and guardrail models with the Apache 2.0 license, according to IBM. IBM states that releases disclose training data sources, methodologies and architectural decisions.
• AI21 Labs: Jamba models with a hybrid Transformer-Mamba architecture, with Jamba2 3B and Jamba2 Mini under Apache 2.0, and Maestro for agents. It offers private deployment in a VPC or on premises.
• Amazon Nova: text, image, video, speech and embedding models accessed through Amazon Bedrock.