In this article
Liquid AI has released LiquidCodec, the ~74M-parameter streaming audio codec behind LFM3-Audio-2B and LFM3-Omni-2B. It compresses 24 kHz speech into 12 discrete token streams at a 12.5 Hz frame rate — one 80 ms frame yields 12 tokens — then decodes predicted token sequences back to waveforms. Weights are available on Hugging Face, with a model card listing 73.6M parameters.
Liquid AI measured the codec in FP32 on a single CPU thread of an Apple M4 Max MacBook Pro. It encodes at about 361× real time and decodes at about 87× real time: encode RTF 0.0028 and decode RTF 0.0115. The same run puts Mimi at 0.0361 encode and 0.0329 decode, NanoCodec at 0.2145 and 0.2543, and X-Codec 2 at 0.7359 and 0.0379. That is roughly 13× lower encode RTF and 2.9× lower decode RTF than Mimi.
LiquidCodec replaces Kyutai's Mimi, which Liquid AI used in LFM2-Audio-1.5B. Mimi's residual vector quantization (RVQ) creates sequential dependencies: each codebook approximates the error left by earlier codebooks. LiquidCodec instead uses grouped finite scalar quantization (GFSQ) inside a symmetric encoder/decoder built from LFM3 blocks. The encoder is fully causal. The decoder predicts spectral coefficients directly, which Liquid AI says avoids learned transposed-convolution upsampling used by comparable waveform decoders.
In GFSQ, features are split into independently quantized groups; each group produces a token from a codebook defined by all possible value combinations, and quantized vectors are concatenated. Because the 12 codebooks are independent, MaskGIT-style transformers can predict them in parallel rather than paying ordered autoregressive steps. Liquid AI argues that makes the token space easier for downstream speech models to learn.
LiquidCodec was pretrained on 264,000 hours of speech across 107 languages with waveform and spectral losses and no adversarial losses, then fine-tuned on 55,000 hours of full-band speech at 24 kHz or higher with adversarial critics. Liquid AI says the second stage removed metallic artifacts left by reconstruction-only pretraining. A final decoder-only quantization-aware training (QAT) step cuts the Q4_0 decoder footprint by 3.5×.
In Liquid AI's reconstruction tests on eight sets — LibriSpeech test-clean, Seed-TTS English and Mandarin, and FLEURS German, French, Mandarin, Japanese, and Korean — LiquidCodec posted the best ViSQOL and lowest mel distance on all eight against X-Codec 2, NanoCodec, and Mimi at 12 codebooks. LibriSpeech ViSQOL drops from 4.319 to 4.292 with Q4_0, still above all three external baselines. In zero-shot TTS with 101,000 hours across six languages and fixed architecture, compute budget, and batch order, Liquid AI reports LiquidCodec tokens reached the highest UTMOSv2 on all eight sets and the best WER/CER on four. LibriSpeech WER was 3.15% versus 6.27% for Mimi; repeating with a Qwen3-1.7B backbone gave consistent results. NanoCodec led PESQ (WB) on seven of eight sets and posted higher LibriSpeech speaker similarity, 0.836 versus LiquidCodec's 0.785.
| Codec | Parameters | Sample rate | Frame rate | Codebooks | Encode RTF* | Decode RTF* | LibriSpeech mel dist.* | LibriSpeech TTS UTMOSv2* |
|---|---|---|---|---|---|---|---|---|
| LiquidCodec | ~74M (73.6M card) | 24 kHz | 12.5 Hz | 12 | 0.0028 | 0.0115 | 0.120 | 3.31 |
| Mimi | 96.2M | 24 kHz | 12.5 Hz | 12 measured (8 at 1.1 kbps design) | 0.0361 | 0.0329 | 0.160 | 2.89 |
| NanoCodec | 62M | 22.05 kHz | 12.5 fps | 13 | 0.2145 | 0.2543 | 0.122 | 3.19 |
| X-Codec 2 | 0.8B | 16 kHz | 50 tokens/s | 1 | 0.7359 | 0.0379 | 0.123 | 3.27 |
*Liquid AI's measurements, FP32 on one Apple M4 Max CPU thread; Mimi run at 12 codebooks. RTF lower is better; mel distance lower is better; UTMOSv2 higher is better.
LiquidCodec loads through Hugging Face Transformers with trust_remote_code=True. The model card shows encode() returning audio codes and decode() returning waveform values, with examples using CUDA and bfloat16. A native LiquidCodecStreamingCache supports chunk-by-chunk decoding and finalization. For edge inference, Liquid AI lists Q4_0, Q4_K_M, and Q8_0 GGUF builds for llama.cpp, shipped alongside LFM3-Audio-2B and LFM3-Omni-2B. The Hugging Face card marks the license as "other," so teams should review the linked LFM License before production use.
AI Mastery analysis
Combined encode and decode RTF is 0.0143, or about 14.3 ms of CPU time per second of audio — roughly 70× real time. That leaves single-thread headroom for language-model inference on hardware matching the Apple M4 Max profile, but Liquid AI has not published x86, lower-clocked Arm, or multi-thread scaling data, so CPU-only deployment needs local benchmarks.
GFSQ's 12 independent codebooks change token scheduling. At 12.5 Hz, serial RVQ codebook prediction would add latency inside every 80 ms frame; parallel MaskGIT-style prediction removes that dependency. But NanoCodec's PESQ lead and its 0.836 LibriSpeech speaker similarity against LiquidCodec's 0.785 mean the speed advantage is not free. Systems prioritizing speaker identity or wideband perceptual score may need compensation downstream. The pattern is component-level design rather than model scaling: architectural specificity outperforms GPU scaling.
Sources
Frequently asked questions
How fast does LiquidCodec encode and decode audio?
Liquid AI reports about 361× real-time encoding and about 87× real-time decoding in FP32 on a single Apple M4 Max CPU thread. The measured RTFs are 0.0028 for encoding and 0.0115 for decoding.
Does LiquidCodec run on a single CPU?
Yes, in Liquid AI's tests it ran on one Apple M4 Max CPU thread. Liquid AI has not published x86, lower-clocked Arm, or multi-thread scaling data, and edge deployment is supported with Q4_0, Q4_K_M, and Q8_0 GGUF builds for llama.cpp.
What changed between LiquidCodec and Mimi?
LiquidCodec replaces Mimi's residual vector quantization with grouped finite scalar quantization. Its 12 GFSQ codebooks are independent, allowing MaskGIT-style parallel prediction instead of sequentially dependent RVQ codebook steps.
How does LiquidCodec compare to NanoCodec?
In Liquid AI's tests, LiquidCodec led on ViSQOL, mel distance, and TTS UTMOSv2 across all eight test sets. NanoCodec led PESQ wideband on seven of eight sets and scored higher LibriSpeech speaker similarity, 0.836 versus 0.785.
Related Reading
Quantize Falcon-7B with LitGPT to Cut GPU Memory to 5.4 GB
Run Falcon-7B locally with LitGPT quantization: exact commands and reference numbers for bf16, NF4, FP4, double quantization, and INT8.
Gradium Phonon 31M Ships 6.9 ms CPU TTS Under Apache 2.0
Gradium's 31M English TTS model runs locally from a 64.5 MB file, reaches 6.9 ms first audio on an M5 CPU, and pairs Apache 2.0 with CC BY-NC multilingual variants.

Fine-Tune a Quantized Llama on GSM8K with PEFT LoftQ
A step-by-step walkthrough of the Hugging Face PEFT GSM8K LoftQ script: 4-bit NF4 loading, preprocessing, training, and exact-match accuracy.