Tsinghua AIR et al.

X-VLA (per-task)↗

X-VLA (per-task) is a VLA from Tsinghua AIR et al., released October 11, 2025. It has no overall rank on the Robot Intelligence Index, which needs both a simulation and a real-robot result. Its best result is #12 of 22 on RoboChallenge. Soft-prompted 0.9B transformer VLA with embodiment-specific prompts for cross-embodiment flow-matching control. (Fine-tuned separately per task.)

Release date: Oct 11, 2025

Type
VLA
Model weights
Open
License
Apache-2.0
Parameters
0.9B
Pricing
Self-hosted
Benchmarks
1
Overall rank
Not ranked
Needs a simulation result to rank overall.
Real robots rank
#32of 56
Rating 1420
Simulation rank
Not ranked
No simulation results yet.
Rating
1398
10601989
Cost / task · est.
$0.201
$0.0087$17.74
Time / task
3min40s
9.5 s1 h 22 min
Rank on each benchmark
BenchmarkScoreRank
RoboChallengeProgress34.812 / 22

Per task attempt; the tiles show the median across benchmarks. Self-hosted compute is priced at one H100 at $3.29 per hour (Lambda on-demand, Oct 2026).

Other models from Tsinghua AIR et al.

Inference

0.9B (24-layer, 1024 hidden). Repo inference example returns 30 predicted 20-D actions per call; no latency published.