36Kr introduces edge-side inference engine, touts reduced latency for large models on robots

eu.36kr.com reported on Sept. 29, 2026, in an article titled “10.7x Faster: This Open-Source Edge-Side Inference Engine Enables Robot Bodies to Run Large Models Without Lag.”

Running large models on robots themselves could be important for applications where communication delays are undesirable, but the original article needs to be checked to determine how fully its assumptions are met. If the announcement is for software, evaluation should focus on the conditions under which its performance claims can be reproduced on actual hardware.

This story stands out for framing improved robot performance as a software-layer issue: how to use computing resources on the robot itself. The “10.7x Faster” wording in the headline indicates that the original article is emphasizing a speed improvement, with the focus apparently on the inference execution method rather than the robot hardware itself. However, only the headline and publication date have been confirmed so far. Without the full original article, it is not possible to determine under what conditions, against what baseline or at which processing stage the improvement was achieved. We have no related coverage to trace continuity with earlier reporting, but publicly available information suggests that as robot inference shifts from reliance on the cloud to processing on the device, latency and communications overhead tend to become key issues. An edge-side inference engine such as this one can be seen as one implementation approach to those issues. In practical use, however, constraints remain around model size, power consumption, heat and onboard hardware, so speed metrics from a laboratory setting need to be considered separately from reliable operation in the field. The key question is not the general integration of AI and robotics, but how much the bottlenecks in running large models on robots themselves can be reduced. At the same time, publicly available information does not confirm the product name, developer, supported hardware, price, licensing terms or benchmark conditions. For now, the technology should be viewed as an indication of the direction of development. Points to check include: 1) the comparison baseline and measurement method; 2) supported robot operating systems and chips; and 3) latency, power consumption and stability on actual hardware. See the original article for details.