MolmoAct 2

Allen Institute for AI (Ai2) · 2026-05

VLANew in 20265Bflow matching

Action reasoning model; the Think variant uses adaptive-depth reasoning tokens.

Paper

Backbone LLM: LLM inside Molmo2-4B (not stated) · Molmo2-ER
Vision encoder: —
Action head: DiT-style flow expert (per-layer KV) (flow matching)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): Molmo2-ER embodied reasoning VLM (starting from Molmo2-4B mid-training checkpoint) + per-layer KV-conditioned DiT-style flow matching action expert (same layer count as backbone); OpenFAST tokenizer

Pretraining accelerators: 64 H100 (Molmo2-ER: 2 nodes × 8 H100)
Pretraining duration: Not disclosed (stated only as GPU-hours)
Pretraining steps & batch: 200K steps, global batch 128, max sequence 4200; Molmo2-ER stage 1 20K + 1.5K steps
GPU/TPU-hours: 5,760 (Stated · H100-hours (about 5,760))
Fine-tuning: GPU-hours not disclosed
Data type: Real-robot teleoperation + web vision-language (embodied reasoning VQA)
Data scale: Post-training set: BimanualYAM 34.5k demonstrations·720+ h, SO-100/101 38,059 episodes·19.8M frames·about 184 h, DROID 74,604 episodes·17,758,044 frames; MolmoER 3.3M samples; total pretraining volume not disclosed (robot 90%·multimodal 10%)
Main datasets: BimanualYAM, SO-100/101, DROID, MolmoER
Tasks & evaluation: LIBERO 97.2% (Think 98.1%); MolmoBot 20.6% (π0.5 10.3%); RoboEval 0.443; real-world zero-shot Franka 87.1% (π0.5 45.2%); Cortex AI 0.51; Molmo2-ER average 63.8 over 13 benchmarks
Robot: Franka, bimanual YAM, SO-100/101

Inference device & speed: LIBERO single H100 default 180 ms, adaptive-depth inference 790 ms (MolmoAct 6,700 ms); bf16 about 12 GB
Inference memory: bf16 about 12 GB (H100)
Weights bf16 (calc.): 10 GB
Where it runs: Not disclosed

Reinforcement learning: Not used · Next-token prediction + flow matching supervised learning
RL evidence: https://arxiv.org/html/2605.02881
Base control: Fixed · No base control (bimanual YAM·Franka·SO100/101 manipulators only)

Key features: Action reasoning model; the Think variant uses adaptive-depth reasoning tokens. Released weights, code, and data, and reports 87.1% zero-shot on a real-world Franka.
Notes: Corrections applied: license Apache 2.0 (README). arXiv v1 2026-05-04, Ai2 blog·README 2026-05-05, pre·post-training code released 2026-06-13. Data hours (h) left empty because total pretraining volume is not disclosed.
Verification: Corrections applied