Physix Frontier · News Briefing Card (Leiphone · Aug 27, 2026)

Leju Launches KUAVO VLA Model, Doubling Industrial Efficiency

KEY FACTS

Leju has introduced the KUAVO VLA model, utilizing over 600 hours of homogeneous real-robot data for mid-training.
The model focuses on common industrial capabilities, achieving a 100% success rate in certain tasks and doubling overall efficiency.
In evaluations across 25 tasks, it ranked first with a 48.27% success rate and a process score of 74.51%.
Compared to the base model Lingbot-VLA 2.0, the new model improves task success rates by 32 percentage points.
It aims to lower the barrier for secondary development by reusing atomic actions to reduce repetitive data collection and debugging.

KEY DATA

600+ hoursMid-training Data Volume

48.27%Overall Task Success Rate

74.51%Process Score

32%Success Rate Improvement vs. Base Model

PHYSIX OBSERVATION

Leju's move marks a shift in embodied AI deployment from 'generalized versatility' to 'vertical specialization.' By using mid-training to bridge the gap between base models and specific skills, this approach significantly cuts industrial deployment costs while addressing developers' pain points of scarce data and cumbersome debugging. Although tying deeply to a single robot body sacrifices cross-model universality, it delivers exceptional execution stability and efficiency, offering the industry a replicable shortcut to commercialization.

Source: Leiphone report