← Leaderboard

DreamZero

World-action
NVIDIA · Feb 17, 2026

World-action model from the paper that introduced the term “World Action Model”, using a video model as a zero-shot policy. It ranks second on RoboArena.

Type
World-action
Parameters
14B
Backbone
Wan2.1-I2V-14B
Weights
Open
License
CC-BY-NC-4.0 (DROID) / Apache-2.0 (AgiBot)
Commercial use
Unclear
Released
Feb 17, 2026
Robots
DROID Franka, AgiBot G1, YAM
SO-101
–
LeRobot
–

Benchmark results

2 of 13 benchmarks
RoboLab-120Simulation
#16 of 2225.7%
RoboArenaReal robot
#2 of 131735
Not reported on 11 benchmarks

Adoption

updated Nov 17, 2026
HF downloads · 30d
2.6K
HF downloads · all time
16.5K
HF likes
39
GitHub stars
2.8K
GitHub forks
248

Running the model

HardwarePrecisionConfigLatency / chunkStepsVRAMChunkControlSource
2× GB200nvfp4DreamZero-Flash, 1 step, all optimizations150 ms1–4830 HzFeb 17, 2026 ↗
GB200–repo inference server with DiT caching600 ms––––Mar 18, 2026 ↗
H100–repo inference server with DiT caching3,000 ms––––Mar 18, 2026 ↗
L40–RoboLab-120 evaluation (DROID)–––24–May 16, 2026 ↗
unspecified–unoptimized baseline5,700 ms4–4830 HzFeb 17, 2026 ↗

Trained on

Runs on

Wrong or missing number? Report an error