Liquid AI Releases LiquidCodec, a 74M-Parameter Streaming Audio Codec

Every voice model has a quiet piece of plumbing underneath it: the codec that turns speech into tokens and tokens back into sound. Liquid AI has now released its own, LiquidCodec, at roughly 74 million parameters, and it is the codec powering its LFM3-Audio-2B and LFM3-Omni-2B models. The weights are on Hugging Face.

The headline number is speed. On a single CPU thread of an Apple M4 Max laptop, Liquid AI reports encoding at about 361 times real time and decoding at about 87 times. If that holds up outside the lab, the codec stops being a bottleneck for voice assistants that run on phones, laptops and edge devices rather than in a data center.

We like this release, and not because of a leaderboard. Voice is where small, local models make the most sense: lower latency, less audio shipped off to someone else's server, and lower cost per conversation. A codec that encodes in a fraction of the time Kyutai's Mimi takes, by Liquid AI's own measurement, pushes that whole category forward.

The design choices are worth a sentence. Instead of residual vector quantization, where each codebook patches what the previous one missed, LiquidCodec uses grouped finite scalar quantization, so its 12 codebooks are independent and can be predicted in parallel. The encoder is fully causal, which suits streaming, and the decoder predicts spectral coefficients directly rather than upsampling through heavy convolution layers. Those are sensible engineering trade-offs aimed squarely at real-time use.

Now the caution. Every benchmark here is self-reported. Liquid AI picked the test sets, the baselines and the hardware. The company deserves credit for showing where it loses: NVIDIA's NanoCodec beat it on PESQ on seven of eight test sets and on speaker similarity on LibriSpeech. Still, I want to see independent reruns before treating the quality lead as settled.

The licensing deserves a hard look too. The Hugging Face card lists the license as "other" and the blog points to Liquid's own LFM License, while Mimi ships under CC-BY 4.0. For builders, that difference can matter more than a few tenths of a quality score.

And there is a downside that comes with any cheaper, better speech pipeline. The same efficiency that makes a private on-device assistant practical also makes convincing synthetic voices cheaper to produce at scale. Codecs are not voice cloning tools on their own, but they are part of the stack, and the industry should be building consent and detection alongside the speedups.

LiquidCodec replaces Mimi, which Liquid AI used in its earlier LFM2-Audio-1.5B model. Training ran in two stages: 264,000 hours of speech across 107 languages, then a fine-tune on 55,000 hours of full-band speech with adversarial critics, which Liquid AI says removed metallic artifacts. Quantized GGUF builds are listed for llama.cpp, and a 4-bit decoder shrinks the footprint 3.5 times with only a small quality dip on LibriSpeech.