Identity
- Status
- active
- Developer ID
allenai/MolmoAct2- Released
- May 5, 2026
- Version
- 2
- License
- apache-2.0
Ai2's open action reasoning model for real-world robot control and post-training across single-arm and bimanual tasks.
0 current hosted offerings
This record is downloadable or self-hosted. Its cost depends on the infrastructure and serving configuration, not a universal model price.
At a Glance
allenai/MolmoAct2Publisher Media
Embodied Intelligence
Open bimanual YAM, SO-100/SO-101, DROID, BC-Z, Fractal, Bridge, and prior MolmoAct data described by Ai2.
Verified Platform Links
| Robot | Relationship | Form | Availability | Evidence note | Source |
|---|---|---|---|---|---|
SO-101Hugging Face LeRobot | Training embodiment | Fixed manipulator | Open build | Ai2 lists SO-100/SO-101 datasets in the MolmoAct 2 training mixture and publishes an SO-100/SO-101 checkpoint. | Official fact |
This page represents one developer model product. Dated API snapshots, serving endpoints, pricing, regions, quantizations, and service tiers attach as versioned aliases or provider details and never create duplicate public model records.
Primary Evidence
Comparable Peers
| A | Pair | B | Context |
|---|---|---|---|
MolmoAct 2Ai2 | vs | Helix 02Figure | cross-developer peersrobotics |
MolmoAct 2Ai2 | vs | Gemini Robotics 2Google DeepMind | cross-developer peersrobotics |
MolmoAct 2Ai2 | vs | SmolVLA 450MHugging Face LeRobot | cross-developer peersrobotics |
MolmoAct 2Ai2 | vs | GR00T N1.7 3BNVIDIA | cross-developer peersrobotics |
MolmoAct 2Ai2 | vs | OpenVLA 7BOpenVLA Research Team | cross-developer peersrobotics |
MolmoAct 2Ai2 | vs | pi 0.7Physical Intelligence | cross-developer peersrobotics |
Questions
Ai2's open action reasoning model for real-world robot control and post-training across single-arm and bimanual tasks. It is developed by Ai2 and its current sourced lifecycle status is active.
MolmoAct 2's current record lists Text, Image, and Robot state as input and Robot action as output.
MolmoAct 2 is classified as a action reasoning model. Its action representation is Flow-matching continuous action expert, its control architecture is Molmo 2-ER backbone with KV-cache bridge and action expert, and inference is recorded as flexible.
The developer reports SO-100, SO-101, Franka, WidowX, and bimanual YAM as supported or represented embodiments. Model Markets also has sourced model-to-platform links for SO-101; each link states whether it is native, demonstrated, used for training, or used for evaluation.
The current record lists 5B total parameters and an unknown active parameter count. Its architecture is Molmo 2-ER backbone with KV-cache bridge and action expert.
MolmoAct 2's recorded capabilities are action-reasoning, bimanual-manipulation, depth-reasoning, and fine-tuning. Its supported tasks are robot-control, dexterous-manipulation, and bimanual-manipulation.
Open bimanual YAM, SO-100/SO-101, DROID, BC-Z, Fractal, Bridge, and prior MolmoAct data described by Ai2.
MolmoAct 2 is explicitly marked as not API-available in the current record.
MolmoAct 2 has no verified hosted API offering, so no provider token price applies. It is a downloadable or self-hosted artifact whose cost depends on the serving infrastructure and workload.
MolmoAct 2 is an open-weight model. Self-hosting is supported, and the recorded license is apache-2.0.