DM0

An “embodiment-native” VLA model open-sourced by Dexmal in 2026, pretrained from the start on a mix of driving and robot data.

DM0 is a vision-language-action (VLA) model released and open-sourced in February 2026 by Dexmal (原力灵机) together with StepFun (阶跃星辰). Most VLA models start from a vision-language model that has only ever seen internet images and text, then fine-tune it on robot data; DM0 instead follows an “embodiment-native” philosophy, training from the start on a mix of web text, autonomous-driving scenes, and robot interaction logs — about 1.2 trillion tokens in all. Training runs in three stages: pretraining a unified VLM (its language component is based on Qwen3-1.7B), mid-training that adds a flow-matching action expert (a sub-network dedicated to producing continuous actions) on top of it, and post-training that fine-tunes with a mixed strategy. When training on embodied data, gradients from the action expert are not backpropagated into the VLM, to preserve its general understanding ability; the model also uses a spatial chain of thought, reasoning about spatial relationships before producing an action. It has about 2 billion parameters and supports both manipulation and navigation in one model. In July 2026, Dexmal released a roughly 6-billion-parameter successor, DM0.5, along with an accompanying open-source framework called OpenDM.

ExampleAccording to the paper, on the RoboChallenge Table30 real-robot benchmark, DM0 reached an average success rate of 62.0% in the task-specific setting and 37.3% in the general setting, ranking first on both at the time of release.

Also called

Dexmal DM0, DM0.5, An Embodied-Native Vision-Language-Action Model towards Physical AI

Related

Vision-Language-Action Model · Action Expert · Flow Matching · Mid-training · RoboChallenge · Dexbotic (Dexmal VLA toolbox)

Sources

DM0: An Embodied-Native Vision-Language-Action Model towards Physical AI (arXiv 2602.14974)
Dexmal/DM05 模型卡（Hugging Face） (Chinese)
Dexbotic 代码仓库（GitHub） (Chinese)

As of

2026-09