MolmoAct 7B-D Pretrain
MolmoAct is a fully open-source action reasoning model for robotic manipulation developed by the Allen Institute for AI. MolmoAct is trained on a subset of OXE and MolmoAct Dataset, a dataset with 10k high-quality trajectories of a single-arm Franka robot performing 93 unique manipulation tasks in both home and tabletop environments. It has state-of-the-art performance among vision-language-action models on multiple benchmarks while being fully open-source. You can find all models in the MolmoAct family here. Learn more about MolmoAct in our announcement blog post or the paper.
MolmoAct 7B-D Pretrain is based on Qwen2.5-7B and uses SigLip2 as the vision backbone, which is initialized using Molmo's pre-training approach. It is pre-trained on MolmoAct's Pre-training Mixture. This model is intended to be used for downstream mid-training, or for replicating our zero-shot results on SimplerEnv (Google Robot).
This checkpoint is a preview of the MolmoAct release. All artifacts used in creating MolmoAct (data, training code, evaluations, intermediate checkpoints) will be made available at a later date, furthering our commitment to open-source AI development and reproducibility.
Update: Checkpoints are now stored in FP32 (previously BF16). The model was trained in FP32, so publishing FP32 weights aligns with training and enables fine-tuning or continued training directly from this repo. For inference, you can still run BF16 by casting at load, which is what we did for evaluations. See more in the instructions below.
Quick links:
Quick Start
To run MolmoAct, first install dependencies:
pip install einops torchvision accelerate
pip install transformers==4.52
Then, follow these steps:
from transformers import AutoProcessor, AutoModelForImageTextToText
import torch
from PIL import Image
import requests
from io import BytesIO
ckpt = "allenai/MolmoAct-7B-D-Pretrain-0812"
# load the processor
processor = AutoProcessor.from_pretrained(
ckpt,
trust_remote_code=True,
torch_dtype="bfloat16",
device_map="auto",
padding_side="left",
)
# load the model
model = AutoModelForImageTextToText.from_pretrained(
ckpt,
trust_remote_code=True,
torch_dtype="bfloat16",
device_map="auto",
)
# task instruction
instruction = "pick orange can"
# strictly follow this reasoning prompt
prompt = (
f"The task is {instruction}. "
"What is the action that the robot should take. "
f"To figure out the action that the robot should take to {instruction}, "
"let's think through it step by step. "
"First, what is the depth map for this image? "
"Second, what is the trajectory of the end effector? "
"Based on the depth map of the image and the trajectory of the end effector, "
"what is the action that the robot should take?"
)
# apply chat template
text = processor.apply_chat_template(
[
{
"role": "user",
"content": [dict(type="text", text=prompt)]
}
],
tokenize=False,
add_generation_prompt=True,
)
# image observation
url = "https://huggingface.co/allenai/MolmoAct-7B-D-Pretrain-0812/resolve/main/example.png"
r = requests.get(url, headers={"User-Agent": "python-requests"}, timeout=30)
r.raise_for_status()
img = Image.open(BytesIO(r.content)).convert("RGB")
imgs = [img]
# process the image and text
inputs = processor(
images=[imgs],
text=text,
padding=True,
return_tensors="pt",
)
# move inputs to the correct device
inputs = {k: v.to(model.device) for k, v in inputs.items()}
# generate output
with torch.inference_mode():
with torch.autocast("cuda", enabled=True, dtype=torch.bfloat16):
generated_ids = model.generate(**inputs, max_new_tokens=256)
# only get generated tokens; decode them to text
generated_tokens = generated_ids[:, inputs['input_ids'].size(1):]
generated_text = processor.batch_decode(generated_tokens, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
# print the generated text
print(f"generated text: {generated_text}")
# >>> The depth map of the image is ... The trajectory of the end effector is ...
# Based on these information, the action that the robot should take is ...
# parse out all depth perception tokens
depth = model.parse_depth(generated_text)
print(f"generated depth perception tokens: {depth}")
# >>> [ "<DEPTH_START><DEPTH_1><DEPTH_2>...<DEPTH_END>" ]
# parse out all visual reasoning traces
trace = model.parse_trace(generated_text)
print(f"generated visual reasoning trace: {trace}")
# >>> [ [[242, 115], [140, 77], [94, 58], [140, 44], [153, 26]]] ]
# parse out all actions, unnormalizing with key of fractal20220817_data
action = model.parse_action(generated_text, unnorm_key="fractal20220817_data")
print(f"generated action: {action}")
# >>> [ [0.0732076061122558, 0.08228153779226191, -0.027760173818644346,
# 0.15932856272248652, -0.09686601126895233, 0.043916773912953344,
# 0.996078431372549] ]
License and Use
This model is licensed under Apache 2.0. It is intended for research and educational use. For more information, please see our Responsible Use Guidelines.
Model and Hardware Safety
MolmoAct offers the ability to inspect a visual trace of its intended actions in space before they occur, allowing users to ensure safe behavior by proactively auditing and adjusting the actions of any hardware acting under the model’s instructions. MolmoAct’s action space is bounded within the data provided, and compliance is built into the model to prevent excessive force when resistance is detected. Please follow the hardware manufacturer’s guidelines when using this model with a robot and perform all operations in a safely configured environment.
Citation
@misc{molmoact2025,
title={MolmoAct: Action Reasoning Models that can Reason in Space},
author={Jason Lee and Jiafei Duan and Haoquan Fang and Yuquan Deng and Shuo Liu and Boyang Li and Bohan Fang and Jieyu Zhang and Yi Ru Wang and Sangho Lee and Winson Han and Wilbert Pumacay and Angelica Wu and Rose Hendrix and Karen Farley and Eli VanderBilt and Ali Farhadi and Dieter Fox and Ranjay Krishna},
year={2025},
eprint={2508.07917},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2508.07917}
}