Liquid AI releases LiquidCodec, a 74M parameter audio codec optimized for on-device speed

Details

Liquid AI released LiquidCodec, a 74M parameter streaming audio codec supporting the LFM3-Audio-2B and LFM3-Omni-2B models, designed to convert audio into tokens predictable by language models.
On an Apple M4 Max MacBook Pro, LiquidCodec achieves approximately 361x real-time encoding and 87x real-time decoding, significantly faster than competitors like Mimi and NanoCodec.
In reconstruction benchmarks across 8 test sets, LiquidCodec recorded the lowest Mel Distance and highest ViSQOL scores, maintaining quality even with Q4_0 quantization which reduces the decoder footprint by 3.5x.
For zero-shot multilingual TTS, models trained on LiquidCodec tokens achieved the highest naturalness scores (UTMOSv2) on all 8 test sets and the highest intelligibility (WER/CER) on 4 of them, with results consistent across LFM3 and Qwen3-1.7B backbones.
The architecture employs Grouped Finite Scalar Quantization (GFSQ) to reduce codebook dependency and enable parallel prediction, replacing the Residual Vector Quantization (RVQ) used in previous models like Mimi.
Training involved two stages: pre-training on 264,000 hours of data across 107 languages, followed by fine-tuning on 55,000 hours of high-fidelity 24 kHz data using adversarial critics to remove artifacts.
The codec is available on Hugging Face and included in GGUF quantized builds (Q4_0, Q4_K_M, Q8_0) for the LFM3 audio models.

Key point