Liquid AI's audio codec encodes 361 times faster than real time

single source· 1 articles · confidence: medium · first seen 2026-11-12 22:43 UTC

What this means for you

Nothing to act on unless you already run LFM3-Audio-2B or LFM3-Omni-2B, where the codec arrives as part of the model; there is no standalone licence or weights link. The number to watch is decoding speed, which is not reported — fast encoding is only half of a live audio pipeline.

Liquid AI released LiquidCodec, a 74M-parameter audio codec — software that turns speech into tokens a model can read, and back again — which the company says encodes at roughly 361 times real time on a single CPU thread. It produces 12.5 frames a second across 12 codebooks (parallel token streams), and it is the codec behind Liquid's LFM3-Audio-2B and LFM3-Omni-2B speech models. Beyond encoding speed the release gives no figures: no decoding cost, no latency, no licence, and no comparison against existing codecs.

Key facts

  • ·LiquidCodec has about 74 million parameters. source
  • ·Liquid AI says it encodes at roughly 361 times real time on a single CPU thread. source
  • ·It runs at 12.5 Hz with 12 codebooks. source
  • ·The codec is used by the LFM3-Audio-2B and LFM3-Omni-2B models. source

What the sources say

  • MarkTechPost — Brief release note carrying the codec's size, frame rate and single-thread encoding speed.

Sources

The original reporting. Follow these — they did the work.

← the wire