MolmoAct2 is an open vision-language-action model for robot control. It builds on Molmo2-ER, an embodied-reasoning VLM backbone, and connects the autoregressive VLM to a flow-matching continuous action expert through per-layer KV (key-value) conditioning. This checkpoint is the post-trained, multi-embodiment MolmoAct2 model. It is intended as a foundation checkpoint for further robot fine-tuning rather than as a ready-to-run policy for a single deployment setting. Use this checkpoint for further fine-tuning on a target robot embodiment or benchmark. It contains the VLM and continuous action expert weights, plus normalization metadata for the post-training mixture in normstats.json. This…
MolmoAct2-Pretrain adapts the Molmo2-ER vision-language backbone into a discrete autoregressive robot policy while keeping the Molmo2 token interface.
Parameters4.9B
Context16,384
Weights19.5 GB
License—
AccessOpen weights
Monthly Downloads18.4k
Runs On
What it takes to serve MolmoAct2-Pretrain (4.9B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 9.7 GB | 11.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 4.9 GB | 5.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 2.4 GB | 2.9 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
Model Card
MolmoAct2-Pretrain adapts the Molmo2-ER vision-language backbone into a discrete autoregressive robot policy while keeping the Molmo2 token interface. Robot state is represented with discrete state tokens, and future one-second actions are represented with OpenFAST action tokens. This checkpoint is the pre-trained VLA backbone before the continuous flow-matching action expert is attached. It is intended for further post-training or fine-tuning, not direct continuous-control inference. Use this checkpoint for further MolmoAct2 training stages. It was converted with addactionexpert=false, so predictaction(...) is intentionally unavailable. Standard Transformers generation can still be used…
Excerpt from the card by Ai2.
Configuration
- Architecture
- MolmoAct2ForConditionalGeneration
- Context length (tokens)
- 16,384
- Layers
- 36
- Hidden size
- 2,560
- Feed-forward size
- 9,728
- Attention heads
- 32
- Key/value heads
- 8
- Head dimension
- 128
- Vocabulary size
- 154,624
- RoPE base
- 5e+06
- Model type
- molmoact2
Identity and Version
- Repository
- allenai/MolmoAct2-Pretrain
- Publisher
- Ai2
- Task
- Robotics
- Modality
- Control
- Library
- transformers
- Parameters
- 4.9B parameters
- Languages
- Not stated by the source
- Revision
- a05effca9ba36c1177359b42a9d5d7a4568dbe3c
- First published
- 2026-05-04
- Last updated
- 2026-05-23
Files and Weights
21 files, 19.5 GB in total. The weights are 4 files totalling 19.5 GB in safetensors.
Weights4 files · 19.5 GB
Configuration11 files · 439.9 KB
Tokenizer2 files · 12.0 MB
Documentation1 file · 3.4 KB
Other2 files · 10.8 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00004.safetensors | Weights | 4.9 GB | c96769f7b5dd |
| model-00002-of-00004.safetensors | Weights | 4.8 GB | 904a2ace082b |
| model-00003-of-00004.safetensors | Weights | 4.8 GB | a90f44933826 |
| model-00004-of-00004.safetensors | Weights | 4.8 GB | d79334a01af2 |
| config.json | Configuration | 3.9 KB | — |
| configuration_molmoact2.py | Configuration | 20.7 KB | — |
| generation_config.json | Configuration | 116 B | — |
| image_processing_molmoact2.py | Configuration | 22.4 KB | — |
| inference.py | Configuration | 25.9 KB | — |
| model.safetensors.index.json | Configuration | 75.9 KB | — |
| modeling_molmoact2.py | Configuration | 191.0 KB | — |
| norm_stats.json | Configuration | 43.8 KB | — |
| processing_molmoact2.py | Configuration | 17.3 KB | — |
| processor_config.json | Configuration | 1.9 KB | — |
| video_processing_molmoact2.py | Configuration | 37.2 KB | — |
| README.md | Documentation | 3.4 KB | — |
| assets/MolmoAct2.svg | Other | 7.6 KB | — |
| chat_template.jinja | Other | 3.2 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer.json | Tokenizer | 12.0 MB | d5395aefc9b1 |
| tokenizer_config.json | Tokenizer | 782 B | — |
License and Download
- License
- Not stated by the source
- Access
- Open weights, no gate
- Download size
- 19.5 GB
Download from Ai2
Released by Ai2 through its official repository on Hugging Face.
Built From
- Described by arXiv:2605.02881
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 19.5 GB |
| 16-bit | 9.7 GB |
| 8-bit | 4.9 GB |
| 4-bit | 2.4 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About MolmoAct2-Pretrain
How much GPU memory does MolmoAct2-Pretrain need?
About 11.7 GB at 16-bit and 2.9 GB at 4-bit: the weights (4.9B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run MolmoAct2-Pretrain on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
What is MolmoAct2-Pretrain's context length?
16,384 tokens, from the maximum position embeddings in its published configuration.
Similar Models
MolmoAct2 is an open vision-language-action model for robot control. It builds on Molmo2-ER and attaches a flow-matching continuous action expert that conditions on the VLM key-value cache through a per-layer connection. This checkpoint is fine-tuned on the SO-100/101 mixture with absolute joint-pose control and annotated language instructions. It is intended for both further fine-tuning and SO-100/101 policy inference. Use this checkpoint for SO-100/101 inference or for further fine-tuning. Dataset normalization metadata is stored in normstats.json. pass normtag="so100so101molmoact2" at inference time. Continuous action prediction is the intended and recommended inference mode. Discrete…
MolmoAct2 is an open vision-language-action model for robot control. It builds on Molmo2-ER and attaches a flow-matching continuous action expert that conditions on the VLM key-value cache through a per-layer connection. This checkpoint is fine-tuned on the full LIBERO training mixture, combining Spatial, Object, Goal, and Long suites. It is intended for both further fine-tuning and LIBERO policy inference. Use this checkpoint for LIBERO inference or for further fine-tuning. Dataset normalization metadata is stored in normstats.json. pass normtag="libero" at inference time. Continuous action prediction is the intended and recommended inference mode. Discrete action prediction is exposed for…
MolmoAct2 is an open vision-language-action model for robot control. It builds on Molmo2-ER and attaches a flow-matching continuous action expert that conditions on the VLM key-value cache through a per-layer connection. This checkpoint is fine-tuned on the bimanual YAM mixture with absolute joint-pose control and annotated language instructions. It is intended for both further fine-tuning and bimanual YAM policy inference. Use this checkpoint for bimanual YAM inference or for further fine-tuning. Dataset normalization metadata is stored in normstats.json. pass normtag="yamdualmolmoact2" at inference time. Continuous action prediction is the intended and recommended inference mode. Discrete…
MolmoAct2-Think extends MolmoAct2 with depth-token reasoning. Before producing an action, the model can predict a compact 10 x 10 discrete depth representation and condition the action expert on the resulting depth-aware VLM cache. This checkpoint is the post-trained, multi-embodiment depth-reasoning model. It is intended as a foundation checkpoint for further robot fine-tuning rather than as a ready-to-run policy for a single deployment setting. Use this checkpoint for further fine-tuning when the downstream policy should use depth reasoning. It contains the VLM, action expert, and depth-token weights, plus normalization metadata for the post-training mixture in normstats.json. This model…
π₀.₅ (Pi05) is a Vision-Language-Action model from Physical Intelligence designed for open-world generalization: it evolves π₀ to generalize to entirely new environments and situations that were never seen during training. The LeRobot implementation is adapted from their open-source OpenPI repository. This policy has been trained and pushed to the Hub using LeRobot. Learn how to train and run it in the LeRobot pi05 guide, or browse the full documentation. The policy consumes these observation features and produces these action features. Inputs Outputs New to LeRobot? These guides cover the full workflow: - Install LeRobot — set up the lerobot package. - Hardware setup — assemble, wire, and…