MLVC: Neural Video Codec Hits 100 FPS Across Apple, Intel, Qualcomm NPUs

Neural video codecs have long outperformed classical standards like HEVC in compression efficiency, but real-world deployment has been stymied by cross-platform inconsistencies and high compute demands. Quantization-based solutions produce non-deterministic results across different hardware, causing catastrophic decoding failures. MLVC, developed by Tanel Pärnamaa and colleagues, tackles this by explicitly transmitting scale parameters through the hyperprior, guaranteeing entropy coding consistency without requiring bit-exact arithmetic. This incurs a slight bitrate overhead, but architectural innovations—gated memory, ReGLU activation, a long-term reference recovery mechanism, and domain-specific perceptual training—recover most of the efficiency. The result is a codec that can run on commodity NPUs from Apple, Intel, and Qualcomm at 100 FPS for both encoder and decoder.

On the VCD video conferencing benchmark, MLVC delivers over 70% BD-rate (MOS) improvement over hardware HEVC, the strongest deployable baseline. Subjectively, it matches or approaches the quality of DCVC-RT, a state-of-the-art neural codec that cannot operate across diverse platforms. This combination of competitive compression, real-time speed, and cross-platform robustness makes MLVC the first neural video codec suitable for widespread deployment on consumer devices. The authors plan to release the code, and the paper has been accepted to ECCV 2026, signaling strong academic validation. For streaming, video conferencing, and cloud gaming, this could be a game-changer.