Allen Institute for AI (Ai2)

Theorizer

Theorizer is an AI system developed by the Allen Institute for AI (Ai2) that synthesizes scientific theories by analyzing large corpora of research literature. It identifies patterns across multiple studies and generates structured claims, facilitating rapid orientation in new scientific domains.

OlmPool v1.0

OlmPool is a controlled suite of 26 fully open 7B language models from Ai2, each released with 38 pretraining checkpoints. It reveals how four architectural choices — QK normalization, grouped-query attention, sliding window attention, and pretraining context length — compound to degrade long-context extension performance by up to 47%.

SERA (Soft-Verified Efficient Repository Agents) v1.0

SERA is an open-source coding agent designed to assist developers in tasks such as code generation, debugging, and maintenance. It offers a cost-effective and efficient method for training coding agents specialized to private codebases, enabling rapid adaptation to specific repositories.

Molmo2-8B v1.0

Molmo2-8B is an open vision-language model developed by the Allen Institute for AI (Ai2) that supports image, video, and multi-image understanding and grounding. Trained on publicly available third-party datasets, it achieves state-of-the-art performance among multimodal models of similar size.

MolmoWeb v2.0

MolmoWeb is an open-source visual web agent developed by the Allen Institute for AI (Ai2). It leverages the Molmo 2 multimodal model family to automate web tasks by interpreting visual interfaces, enabling actions like clicking, typing, and scrolling based on task instructions and live webpage screenshots.

Olmo-3.1 32B Instruct v3.1

Olmo-3.1 32B Instruct is a powerful open-language model from the Olmo 3 family, specifically designed for conversational and reasoning tasks. It is pre-trained on the Dolma 3 dataset and post-trained on Dolci datasets, offering high-level performance across math, coding, and general reasoning benchmarks.

Olmo-3.1 32B Think v3.1

Olmo-3.1 32B Think is a high-performance 32-billion-parameter language model specialized for complex reasoning, multi-step logic, and advanced chain-of-thought processing. As an open-science model, it provides full transparency into its training data, code, and methodology, making it a powerful resource for researchers and developers in fields like mathematics and programming.