Here's what published — published is not granted. Application US20250036920A1, "Mixture-of-Experts Model Implementation Method and System, Electronic Device, and Storage Medium," published January 30, 2025, assigned to Beijing Baidu Netcom Science Technology Co., Ltd., inventors including Liang Shen and Haifeng Wang. The CPC codes are G06N 3/045 and G06N 3/0495 (architectures and model compression).
The mechanism is implementation, not invention of the concept. Mixture-of-experts is, by 2025, a well-established architecture; the contested ground is making it work efficiently in practice — how experts are stored, how tokens are routed and dispatched, how the system scales across devices. This application claims a particular implementation method and system for an MoE model, which is where the real engineering value sits once the architecture is commodified.
The strategic signal is geographic. Baidu's presence in the MoE record alongside Intel, Microsoft, and NVIDIA shows this is a global contest, not a US-only one. Chinese labs file extensively in exactly the same hot subfields, and a freedom-to-operate analysis that ignores non-US assignees misses a large part of the landscape.
Because this is a publication, the discipline holds: claims as filed, not allowed. Scope is undetermined until a grant issues, and even then it would cover this specific implementation, not MoE in general. The compression CPC suggests efficiency is part of the claimed contribution.
The takeaway: US20250036920A1 is a reminder that the MoE patent race is international — Baidu staking implementation claims in the same window as the US incumbents — and that the published-versus-granted and US-versus-global distinctions both matter for any serious ownership read.