A

Allen Institute for AI

Research instituteAI model developer

United States · seattle · Established 2014

Overview

Model families
2
Versions
4
Downloadable models
2
Latest model release
May 5, 2026
RoboAtlas Summary

Ai2 is a Seattle-based nonprofit AI research institute founded by Paul Allen in 2014. It develops foundational research and innovation across large-scale open models, data, robotics and conservation to support real-world applications.

Related models

Publicly sourced records only; these are not an exhaustive list of supported robots or evidence of factory installation. Evidence levels describe public documentation, not model capability.

In-house models · 2

In-house models
Molmo2-ERTypeEmbodied reasoningCurrent versionOpenness
Code and weights
Related robots
None recorded
Latest release
Awaiting verification
MolmoAct2TypeAction policyCurrent versionOpenness
Code and weights
Related robotsLatest release

Model participation

Timeline

  1. Model released

    MolmoAct2 training code released

    Ai2 publishes pre-training and post-training code.

  2. Model released

    MolmoAct2 announces code/data update and official LeRobot integration

    The May28 dated update in Ai2’s May5 post announces code and training data, including fine-tuning scripts, evaluation rollouts and tokenizer training recipe. The project README separately dates the official Hugging Face LeRobot integration to May28. This is an announced ecosystem/data milestone, not proof that every pretraining component first became public that day.

  3. Model released

    Ai2 announces MolmoAct2 action-reasoning models and artifacts

    On May5,2026 Ai2 announces MolmoAct2 with model weights, datasets and an updated action-reasoning architecture: Molmo2-ER, a KV-conditioned continuous flow-matching expert, and a Think depth-reasoning variant. The release includes a720-hour bimanual YAM dataset. Foundation Base/Think checkpoints are for further fine-tuning; dedicated fine-tuned robot policies are separate.

  4. Model released

    Ai2 introduces Molmo2-ER as the MolmoAct2 reasoning backbone

    The May5 Ai2 post identifies Molmo2-ER, a separately specialized vision-language backbone for spatial questions, pointing and image/video reasoning that feeds MolmoAct2’s action model. It reports a63.8/100 mean across13 embodied reasoning benchmarks. This is an introduction date, not a verified first-public-weight day.

References

Source category does not verify the claims it contains.

Submit a source or correction