MolmoSpaces
Ai2Maintainer ↗Pick, pick and place, open, close, randomized camera and commonsense tasks.
As of November 17, 2026, MagicBrain + Qwen3.7-Plus leads with 94.4, followed by MolmoBot-Img (92.8) and MolmoBot (92.0). Official leaderboard ↗ · Paper ↗
Leaderboard
- Tasks
- 19
- Attempts / task
- 1
- Median cost / task
- $0.018
- Median time / task
- 20 s
Success at any point in the episode (%). Blank means the model did not run that task.
MolmoBot92.0———$0.015est.16 sest.
Awomo-0.5V86.458.962.486.0$0.015est.17 sest.
Awomo-0.5I84.571.049.387.2$0.016est.17 sest.
Cosmos3-Nano-Policy65.535.232.079.1$0.018est.20 sest.
MolmoAct248.323.311.771.3$0.021est.23 sest.
π0.536.413.622.765.1$0.022est.24 sest.
π0-FAST22.010.911.138.6$0.024est.26 sest.29
π016.212.511.053.1$0.025est.27 sest.
DreamZero———————
X-VLA——————Tasks
19 tasks. Hover a task to play its clip (8 of 19).
▶ Clip
▶ Clip

▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip







Methodology
- Scoring
- Code check
- Time cap / attempt
- 30 s
- Control rate
- 15.2 Hz
- Models in index
- 8 of 36
Scoring. Simulator pass/fail check, e.g. object lifted 1 cm and held only by the robot.
Who runs it. Ai2 plus submitted results.
Robots. Simulated Franka FR3.
Note. Its combined score counts tasks a model skipped as zero, so we use single tasks.
Time per attempt. Not published. Estimated otherwise: time cap 30 s × (1 − success/2).
Cost per attempt. Not published. Estimated otherwise: one H100 at $3.29 per hour (Lambda on-demand, Oct 2026) × time. GPU price source ↗