Product Launch read

Liquid AI details LiquidCodec, a 74M-parameter speech codec

Liquid AI published details of LiquidCodec, the speech tokenizer behind its LFM3 audio models. The company says it encodes about 13 times faster than Kyutai's Mimi on a single CPU thread while matching or beating it on reconstruction quality.

AE
Aidose Editorial
Nov 14, 2026

Liquid AI has published details of LiquidCodec, the in-house streaming codec that converts speech into tokens for its language models and turns those tokens back into audio. The company described it in a blog post published November 12. It already powers the LFM3-Audio-2B and LFM3-Omni-2B models Liquid released at the end of October.

LiquidCodec has about 74 million parameters, compared with about 96 million for Kyutai's Mimi codec. It is built on Liquid's LFM3 architecture with a symmetric encoder and decoder and a fully causal encoder, which lets it work on streaming audio. It uses grouped finite scalar quantization rather than the residual vector quantization Mimi uses. It runs at 12.5 frames per second, one frame every 80 milliseconds, on 24 kHz audio.

On a single CPU thread on an Apple M4 Max MacBook Pro, Liquid reports encoding at about 361 times real time and decoding at about 87 times real time. Comparing the published real-time factors, that makes encoding about 13 times faster and decoding about 2.9 times faster than Mimi with 12 codebooks.

Liquid says LiquidCodec had the best ViSQOL score and the lowest mel distance on all eight of its test sets, including 4.319 ViSQOL and 0.120 mel distance on LibriSpeech test-clean. Text-to-speech models trained on its tokens had the highest UTMOSv2 naturalness on all eight sets and the best word or character error rate on four of them. Quantization-aware training on the decoder shrinks it by 3.5 times at Q4_0 with minimal quality loss, according to the company.

The codec loads in Hugging Face Transformers, and GGUF builds of the LFM3 audio models run on llama.cpp. For teams building voice agents on CPUs or phones, a codec this fast takes little of the compute budget on the device. The blog post does not state a license for the codec, so developers should check the model page terms before shipping it.

AE
Aidose Editorial
Nov 14, 2026 · read
Back to News