AI2

MolmoAct-7B-O

Allen Institute for AI, USA · August 2025

Allen Institute for AI (Ai2)

Reasoning VLAOpen weights, Apache-2.0Also written MolmoAct-7B-O-0812
Parameters
7.7 B
The allenai/MolmoAct-7B-O-0812 checkpoint reports 7,656,630,272 parameters in its safetensors index (Hugging Face model API,...
GPU memory
14 GB
weights at bf16, computed. Fits a 24 GB card
Inference latency
not published
per action step
Weights
Apache-2.0
downloadable checkpoint

What it is

MolmoAct-7B-O is the maximally open variant of MolmoAct, released alongside MolmoAct-7B-D on the same date. It swaps the Qwen2.5-7B and SigLIP2 stack for an OLMo-2-1124-7B language model with an OpenAI CLIP ViT-L/14-336 vision encoder, so that the language model itself is open weights, open data and open code rather than a third-party release. The paper describes it as its most open model and notes the caveat that the CLIP vision encoder is not itself fully open. Functionally it is the same Action Reasoning Model: depth perception tokens, editable trajectory traces, then discrete action tokens. The important limitation for anyone comparing models is that the paper reports 7B-O only in the architecture and hyperparameter appendices, and publishes no benchmark results for it, so its accuracy relative to 7B-D is undocumented.

Architecture

Backbone
OLMo-2-1124-7B language model with OpenAI CLIP ViT-L/14-336 vision encoder
Action head
Autoregressive discrete action tokens, same three-stage Action Reasoning Model pipeline as MolmoAct-7B-D (depth perception tokens, editable trajectory traces, low-level actions), 256 action bins.
Action chunk
8 steps per forward pass
Parameters
The allenai/MolmoAct-7B-O-0812 checkpoint reports 7,656,630,272 parameters in its safetensors index (Hugging Face model API, checked 11.08.2026). Base models declared on the card: allenai/OLMo-2-1124-7B and openai/clip-vit-large-patch14-336.
Pretraining data
Same MolmoAct Pre-training Mixture, MolmoAct Dataset mid-training and Open X-Embodiment subset as MolmoAct-7B-D. The paper documents separate hyperparameter columns for 7B-D and 7B-O across pre-training, mid-training and post-training.
Embodiments
Not separately documented; the MolmoAct evaluation embodiments (Franka Emika Panda, Google Robot in SimplerEnv, LIBERO) are reported for 7B-D only

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 7.7 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3229 GBTraining master weights, rarely used for inference
bf16 or fp1614 GBThe usual way these checkpoints are served
int87.1 GBQuantised, expect some loss of precision on fine motions
int43.6 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Not published separately from MolmoAct-7B-D. The paper notes a tokenizer detail specific to 7B-O: because OLMo-2-7B has fewer than 130 padding tokens, the tokenizer and lm head must be padded first.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

No real-world results are published for this variant. All real-world numbers in the MolmoAct paper refer to MolmoAct-7B-D.

Fine tuning it yourself

Same pipeline as MolmoAct-7B-D: LeRobot-format datasets preprocessed through the action reasoning pipeline, action chunk size 8, 256 action bins. The repository recommends MolmoAct-7B-D-0812 rather than 7B-O as the checkpoint for fine-tuning on custom datasets.

Where it helps, where it does not

Strengths

  • Most open variant of MolmoAct: the language model (OLMo-2) is open weights, open data and open code, not just open weights
  • Apache-2.0, so usable commercially
  • Useful for reproducibility research where the full data provenance of the language backbone matters

Limits

  • No benchmark results are published for this variant anywhere; the paper mentions 7B-O only in architecture and hyperparameter tables
  • Openness is not complete: the OpenAI CLIP ViT-L/14-336 vision encoder is not fully open, which the paper acknowledges
  • Not the recommended fine-tuning starting point; the repository points users to MolmoAct-7B-D-0812
  • Inherits the MolmoAct generation's slow inference and is superseded by MolmoAct2 (May 2026)
  • No latency, VRAM or control frequency figures

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.