74M Parameters, 361x Real-Time Performance: Liquid AI Turns Audio Codec Into the Foundation of Speech Models

Liquid AI has launched LiquidCodec, a streaming audio codec with roughly 74 million parameters. The codec is integrated into both LFM3-Audio-2B and LFM3-Omni-2B. As reported by Asif Razzaq, writer at MarkTechPost, it runs at 12.5 Hz with 12 codebooks, and achieves an encoding speed of approximately 361 times real-time on a single CPU thread.

To fill in the full context: an audio codec acts as a "translator" for speech models. It compresses continuous audio waveforms into discrete tokens, and after the model processes these tokens, it converts them back into audible audio. Regular end users notice whether the voice chat feels responsive enough, while developers focus on frame rate, number of codebooks, computing power consumption and concurrency capacity. 12.5 Hz multiplied by 12 codebooks equals 150 tokens per second per audio stream. This magnitude determines how many audio tokens need to be stored in the model's context, and also directly affects inference costs and response speed.

The 361x real-time performance figure is even more straightforward. Calculated on a single CPU thread, the theoretical time required to process 60 seconds of audio is only around 0.17 seconds. This does not guarantee optimal audio quality or end-to-end latency, but it proves the computational pressure of the encoding stage can be reduced to an extremely low level. In real-time voice products, this link often goes unnoticed, yet it can easily become a bottleneck that limits maximum concurrency and pushes up operating costs.

My assessment: When large model manufacturers are competing to expand parameter scales, audio codecs are often treated as a secondary supporting component. But in real-time speech scenarios, this "supporting role" directly determines latency. With only 74 million parameters, LiquidCodec is far from eye-catching under the common narrative that hypes large LLM parameter sizes; but in the full audio processing pipeline, it acts as the critical entry point for the model. By pairing LiquidCodec tightly with LFM3-Audio-2B and LFM3-Omni-2B, Liquid AI has made its intention very clear: beyond core model capabilities, it wants full control over the underlying supporting infrastructure as well.

Source: MarkTechPost, author Asif Razzaq, Original Article Link, published on November 12, 2026. No official product page or academic paper is attached to the published materials, so public disclosures are still needed to confirm metrics including audio quality, latency and bitrate.