Microsoft open-sources MLVC neural codec to solve cross-platform drift

Microsoft’s transition of MLVC from lab research to production testing in Teams marks the first major deployment of a neural codec in a mainstream RTC platform. By leveraging general-purpose NPUs, Microsoft bypasses the years-long hardware adoption cycle required by fixed-function silicon like AV1 or VVC. However, the lack of standardized IP pools means adopters face a fragmented landscape of individual patent negotiations rather than the one-stop licensing models of traditional codecs. Watch for the delivery of native C++ SDK bindings, which will signal when third-party RTC platforms can realistically integrate the technology beyond controlled peer-to-peer environments.

The release of MLVC occurs as the standards community intensifies its focus on the 'Beyond VVC' (H.267) era. Per GreyB (March 2026), over 3,900 patents and 1,300 JVET contributions are already shaping two primary tracks: the Enhanced Compression Model (ECM) and Neural Network Video Coding (NNVC). While Microsoft offers MLVC under the MIT license, commercial licensors like Bytedance and Qualcomm are aggressively filing patents to secure standard-essential positions. Qualcomm’s US Patent 11,388,416, for instance, already covers foundational auto-encoder and latent code space methodologies similar to those used in modern neural codecs.

In the broader ecosystem, patent pool administrators are consolidating to address this fragmentation. Access Advance (January 2026) recently integrated Via LA’s HEVC and VVC assets into a renamed 'VCL Advance' program to streamline licensing. Meanwhile, InterDigital’s October 2025 acquisition of AI-native startup Deep Render indicates that established licensors are moving to lock up proprietary neural compression IP. These moves suggest that the clean-slate IP promise of AI codecs is fading; per Streaming Media (March 2026), the acquisition signals that AI-native compression will likely follow the same royalty-heavy trajectories as its predecessors, with full standardization not expected until at least 2028.