Microsoft open sources NPU-optimized neural codec to slash 360p bandwidth
Microsoft has open sourced MLVC, a neural network-based video codec designed to leverage Neural Processing Units (NPUs) for efficient video compression. The codec demonstrates significant bandwidth reductions compared to H.264 and is currently being deployed within Microsoft Teams.
Key Takeaways
- MLVC requires approximately 122 kbps for 360p video at 30 fps, compared to the 1 Mbps typically required by H.264.
- Integrated support is available for Apple Neural Engine, Qualcomm Hexagon NPUs, and Intel OpenVINO, with AMD NPU support pending.
- The codec uses no more than 50% of a chip's native NPU capability to preserve overhead for other concurrent AI tasks.
- Microsoft released the model source code, training weights, and platform conversion scripts under the MIT License.
Why It Matters
Microsoft’s push to productize neural codecs shifts the industry toward end-to-end learned compression, bypassing the slow hardware cycles of traditional standards. By targeting NPUs rather than power-hungry GPUs, Microsoft enables high-efficiency video on mobile and laptop hardware without compromising battery life. For the streaming ecosystem, this indicates a move away from static algorithmic rules toward dynamic, data-driven prediction models. To compete, other cloud and enterprise video providers must now decide whether to adopt Microsoft’s open-sourced weights or develop proprietary neural models. Watch for the standardization of Neural Network Video Coding (NNVC) within the broader "Beyond VVC" initiatives through 2028.
Additional Context
The release of MLVC aligns with a broader industry shift toward AI-native compression, as specialized hardware becomes a baseline for consumer electronics. Per internal Intel reporting from March 2026, the company’s Series 3 'Panther Lake' processors now lead the x86 market with NPU performance reaching 50 TOPS, designed specifically for continuous inference tasks like real-time video decoding. This hardware surge is matched by rival activity; per industry filings in October 2025, InterDigital acquired AI compression startup Deep Render, signaling aggressive consolidation in the neural codec space as traditional standards like VVC face fragmented patent landscapes and slow adoption. While hardware capabilities have matured, the software ecosystem remains the primary bottleneck for neural codec deployment. According to the Alliance for Open Media (AOMedia) in late 2025, the development of AV2 is incorporating similar learned compression techniques to maintain competitiveness against proprietary neural solutions. Current benchmarks from early 2026 suggest that while GPUs remain superior for high-throughput training, NPUs are roughly 40% more power-efficient for the sustained, low-latency inference required for video conferencing and mobile streaming. Microsoft’s open-source move effectively attempts to standardize its framework before competing standards from AOMedia or Apple’s reported PICO image codec capture the developer ecosystem.
Read full article at techstrong.ai
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source