Liquid AI has released LiquidCodec, a 73.6M-parameter streaming audio codec operating at 24 kHz with a 12.5 Hz frame rate and 12 codebooks, designed to replace Kyutai's Mimi in the LFM3-Audio-2B and LFM3-Omni-2B models. The architecture employs symmetric encoder-decoder blocks derived from LFM3, featuring a fully causal encoder suited for streaming. It utilizes Grouped Finite Scalar Quantization (GFSQ) instead of Residual Vector Quantization (RVQ), enabling parallel prediction of all 12 codebooks by MaskGIT-style transformers, while the decoder directly predicts spectral coefficients to avoid costly transposed-convolution upsampling. Training occurred in two stages: 264,000 hours of multilingual speech with waveform and spectral losses (no adversarial loss), followed by fine-tuning on 55,000 hours of full-band speech with adversarial critics to eliminate metallic artifacts, concluding with quantization-aware training (QAT) for edge deployment. On an Apple M4 Max CPU (FP32), LiquidCodec achieves ~361x real-time encoding (RTF 0.0028) and ~87x decoding (RTF 0.0115), outperforming Mimi (0.0361/0.0329), NanoCodec (0.2145/0.2543), and X-Codec 2 (0.7359/0.0379). Across eight test sets (LibriSpeech, Seed-TTS, FLEURS), it achieves the best ViSQOL (4.319 on LibriSpeech) and lowest mel distance (0.120), plus the highest TTS UTMOSv2 (3.31) and best WER/CER on four sets (LibriSpeech WER 3.15% vs Mimi's 6.27%). NanoCodec leads PESQ-WB on seven sets and speaker similarity (SECS 0.836 vs 0.785). Q4_0 GGUF quantization reduces decoder size 3.5x with minimal ViSQOL degradation (4.319 to 4.292). Weights are available on Hugging Face with Transformers and llama.cpp support.