MolmoAct2: Action Reasoning Models for Real-world Deployment
:::info Stub — Full Engineering Breakdown Coming This paper was featured on Hugging Face Daily Papers on 2026-05-04 with 269 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Haoquan Fang et al. |
| Year | 2026 |
| HF Upvotes | 269 |
| arXiv | 2605.02881 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Vision-Language-Action (VLA) models aim to provide a single generalist controller for robots, but today's systems fall short on the criteria that matter for real-world deployment. Frontier models are closed, open-weight alternatives are tied to expensive hardware, reasoning-augmented policies pay prohibitive latency for their grounding, and fine-tuned success rates remain below the threshold for dependable use. We present MolmoAct2, a fully open action reasoning model built for practical deployment, advancing its predecessor along five axes. We introduce MolmoER, a VLM backbone specialized for spatial and embodied reasoning, trained on a 3.3M-sample corpus with a specialize-then-rehearse recipe. We release three new datasets spanning low-to-medium cost platforms, including MolmoAct2-BimanualYAM, 720 hours of teleoperated bimanual trajectories that constitute the largest open bimanual dataset to date, together with quality-filtered Franka (DROID) and SO100/101 subsets. We provide OpenFAST, an open-weight, open-data action tokenizer trained on millions of trajectories across five embodiments. We redesign the architecture to graft a flow-matching continuous-action expert onto a discrete-token VLM via per-layer KV-cache conditioning. Finally, we propose MolmoThink, an adaptive-depth reasoning variant that re-predicts depth tokens only for scene regions that change between timesteps, retaining geometric grounding at a fraction of prior latency. In the most extensive empirical study of any open VLA to date, spanning 7 simulation and real-world benchmarks, MolmoAct2 outperforms strong baselines including Pi-05, while MolmoER surpasses GPT-5 and Gemini Robotics ER-1.5 across 13 embodied-reasoning benchmarks. We release model weights, training code, and complete training data. Project page: https://allenai.org/blog/molmoact2
Engineering Breakdown
Plain English
MolmoAct2 is an open-weight Vision-Language-Action model designed specifically for real robot deployment, addressing key limitations of existing systems like closed-source models, hardware dependencies, and latency overhead. The team built MolmoER, a specialized vision-language backbone trained on 3.3M embodied reasoning examples using a two-stage specialize-then-rehearse approach, and released three new datasets for low-to-medium cost robot platforms including bimanual systems.
Key Engineering Insight
The paper's core engineering win is the specialize-then-rehearse training recipe: pre-training a VLM backbone on spatial and embodied reasoning tasks before fine-tuning on downstream robot control. This design choice directly reduces latency overhead and hardware requirements compared to reasoning-augmented policies, making deployment feasible on commodity hardware.
Why It Matters for Engineers
Most VLA systems today either require proprietary APIs (no production control), demand expensive GPUs, or add unacceptable latency for real-time robot control. MolmoAct2 being fully open-weight and built for low-to-medium cost platforms means teams can actually deploy and iterate on robot systems without vendor lock-in or prohibitive infrastructure costs. The explicit focus on success rate thresholds suggests they're optimizing for reliability metrics that matter operationally.
Research Context
Prior VLA work traded off between generalization (frontier closed models) and accessibility (open models stuck on expensive hardware or slow inference). MolmoAct2 advances the field by proving you can build an open, efficient, reasoning-capable action model at scale. This enables a new class of deployable robot systems and shifts the baseline from "research prototype" to "production-ready" for embodied AI applications.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::