AI2

MolmoAct2-Think

Allen Institute for AI, USA · May 2026

Allen Institute for AI (Ai2)

Reasoning VLAOpen weights, Apache-2.0Also written MolmoAct 2-Think, allenai/MolmoAct2-Think
Parameters
5.5 B
The allenai/MolmoAct2-Think checkpoint reports 5,490,552,304 parameters in its safetensors index, and the Hugging Face...
GPU memory
10 GB
weights at bf16, computed. Fits a 16 GB card
Inference latency
790 ms
Single NVIDIA H100, LIBERO, action horizon 10, with CUDA Graph replay enabled
Weights
Apache-2.0
downloadable checkpoint

What it is

MolmoAct2-Think is the depth-reasoning variant of MolmoAct2, which restores the explicit spatial reasoning stage that the base model drops for speed. It performs adaptive depth reasoning by autoregressively predicting only the depth tokens that changed, so inference cost scales with the dynamic fraction of the scene rather than the whole frame while the geometric grounding is retained. This buys accuracy: 98.1 percent LIBERO average against 97.2 percent for the base model, with the largest gain on LIBERO-Long (95.4 versus 93.2). The cost is latency, since it reaches 12.71 Hz on a single H100 at action horizon 10 against 55.79 Hz for the base model, roughly 790 ms versus 180 ms per action. It also benefits far less from CUDA Graph replay (1.58 times versus 2.42 times) because autoregressive decoding has sequential dependencies and variable-length execution that graph capture cannot absorb.

Architecture

Backbone
Molmo2-ER, an embodied-reasoning VLM built on Molmo2-4B
Action head
Same DiT-style flow matching action expert with per-layer KV conditioning as MolmoAct2, preceded by an adaptive depth reasoning stage that autoregressively predicts only the tokens for changed depth cells and replays cached cells for the static scene fraction.
Action chunk
10 steps per forward pass
Parameters
The allenai/MolmoAct2-Think checkpoint reports 5,490,552,304 parameters in its safetensors index, and the Hugging Face collection labels it 5B (checked 11.08.2026). Marginally larger than the base MolmoAct2 at 5,485,309,424.
Pretraining data
Same 3.3M-sample embodied reasoning corpus and action datasets as MolmoAct2. The Think checkpoint is fine-tuned from the depth-aware post-trained checkpoint using the same full-LIBERO recipe and compute budget as the base model.
Embodiments
Franka Panda in LIBERO simulation (MolmoAct2-Think-LIBERO), Shares the MolmoAct2 embodiment support: bimanual YAM, Franka DROID, SO-100/101

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 5.5 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
fp3220 GBTraining master weights, rarely used for inference
bf16 or fp1610 GBThe usual way these checkpoints are served
int85.1 GBQuantised, expect some loss of precision on fine motions
int42.6 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 16 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.

Published figures. These come from the model authors, not from this site.
Inference latency, reported790 ms on Single NVIDIA H100, LIBERO, action horizon 10, with CUDA Graph replay enabled source
Checkpoint download22 GB

No separate VRAM figure is published for the Think variant. The repository's general guidance applies: checkpoints are about 22 GB each and bfloat16 is the documented safe default. Measurements were taken on a single H100.

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

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • LIBERO Spatial
    98.8%
    success ratePaper Table 8. Ablation shows the full recipe (mixed training plus noise injection plus depth gate) lifts the average from 97.50 to 98.10 percent.source
  • LIBERO Average over 4 suites
    98.1%
    success ratePaper Table 8, best in table. Per split: Spatial 98.8, Object 99.8, Goal 98.5, Long 95.4. Improves over base MolmoAct2 on three of four suites.source
  • LIBERO Long
    95.4%
    success ratePaper Table 8. Largest gain over base MolmoAct2 (93.2), which is where the depth reasoning stage pays off.source

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 separate real-world numbers are published for the Think variant. The paper's real-world zero-shot evaluations report MolmoAct2-DROID and MolmoAct2-SO100/101, which are base-model derivatives.

Fine tuning it yourself

MolmoAct2-Think-LIBERO is fine-tuned from the depth-aware post-trained checkpoint using the same full-LIBERO mixture and compute budget as the base model. The best recipe combines mixed training, noise injection and a depth gate, which the ablation shows lifts the LIBERO average from 97.50 to 98.10 percent. Same LeRobot v3.0 data format as MolmoAct2.

Where it helps, where it does not

Strengths

  • Best LIBERO average of any model in the MolmoAct2 comparison at 98.1 percent, and the best LIBERO-Long at 95.4 percent
  • Adaptive depth reasoning ties inference cost to the dynamic fraction of the scene instead of the whole frame, so the reasoning overhead is not constant
  • Retains the geometric grounding that the base model gives up, while still being roughly 8.5 times faster than the original MolmoAct
  • Apache-2.0 with open weights and code
  • Improves over the base model on three of four LIBERO suites and matches it on the fourth

Limits

  • Roughly 4.4 times slower than base MolmoAct2, about 790 ms versus 180 ms per action on the same H100 setup
  • Benefits much less from CUDA Graph replay (1.58 times versus 2.42 times) because autoregressive decoding is not amenable to graph capture
  • No real-world evaluation results are published for this variant; all real robot numbers belong to base-model derivatives
  • No separate VRAM measurement is published
  • The accuracy gain over the base model is 0.9 points on the LIBERO average, which is small relative to the latency cost for most deployments

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.