Introduction

Liquid AI released LiquidCodec on November 12. It's the in-house audio codec behind its on-device voice models, LFM3-Audio-2B and LFM3-Omni-2B. A codec is the part of a speech model that turns audio into tokens a language model can predict, then turns the predicted tokens back into sound. That makes it the hidden layer that decides how a TTS or voice-chat model sounds and how fast it can talk back.

The weights are open on Hugging Face at LiquidAI/Liquid-Codec. It has about 74M parameters, works on 24 kHz mono speech, emits 12 tokens per 80 ms frame (about 1.65 kbps), and streams in both directions. Liquid reports ~361× real-time encoding and ~87× real-time decoding on a single Apple M4 Max CPU thread. The license is the LFM Open License v1.0.

Sources:

(November 12, 2026), the LiquidCodec engineering blog post and the Hugging Face model card.

What Shipped

  • Name and maker: LiquidCodec, by Liquid AI.
  • Modality: a speech audio codec (tokenizer plus detokenizer). It doesn't speak by itself. It's the audio layer that a TTS, voice-agent or speech-to-speech model builds on.
  • Size: about 74M parameters. The encoder and decoder are symmetric, both built from LFM3 blocks, and the encoder is fully causal for streaming.
  • Tokens: grouped finite scalar quantization (GFSQ) instead of the residual vector quantization (RVQ) that Mimi uses. The 12 codebooks don't chain off each other, so a speech model can predict all 12 tokens of a frame in parallel instead of one after another.
  • Format: 24 kHz mono in and out, 12.5 Hz frame rate, 12 tokens per frame, about 1.65 kbps.
  • Decoder trick: the decoder predicts spectral coefficients directly, which skips the slow learned upsampling stack many codecs use.
  • Training: 264,000 hours of speech across 107 languages, then a second stage on 55,000 hours of cleaner full-band speech with adversarial critics to remove "metallic" artifacts.
  • Edge build: quantization-aware training on the decoder lets it run in Q4_0 at a 3.5× smaller footprint. Liquid says quality loss is minimal.
  • Open vs closed: open weights (BF16 safetensors) plus the model code, under the LFM Open License v1.0.

Where to Run It Today

  • Hugging Face + Transformers: the model card ships remote code, so it loads with AutoModel and trust_remote_code=True on a recent Transformers 5.x (pip install -U transformers torch).
  • On-device, inside the LFM3 models: quantized decoders (Q4_0, Q4_K_M, Q8_0) aren't published separately. They come bundled in LFM3-Audio-2B-GGUF for llama.cpp.
  • Talk to it: Liquid's playground runs the LFM3 audio models that use this codec.

A paste-ready round trip, adapted from the model card. Use device_map="cpu" if you don't have a CUDA GPU:

import torch
from transformers import AutoModel

codec = AutoModel.from_pretrained(
    "LiquidAI/Liquid-Codec",
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map="cuda",
)

waveform = torch.randn(1, 24_000 * 3, device="cuda")  # 3 s of 24 kHz mono audio
codes = codec.encode(waveform).audio_codes            # shape (1, 12, 38): 12 codebooks x 38 frames
reconstructed = codec.decode(codes).audio_values      # back to a waveform

Swap the random tensor for a real 24 kHz mono clip (for example, loaded and resampled with torchaudio) to hear the reconstruction. The repo's model code also includes LiquidCodecStreamingCache for streaming decode (pass it to codec.decode frame by frame with finalize=False, then once with finalize=True), which returns each chunk of audio as soon as later frames can no longer change it. That's the piece that matters for low-latency voice agents.

Numbers (Self-Reported)

Liquid compared LiquidCodec with X-Codec 2, NVIDIA's NanoCodec and Kyutai's Mimi (at 12 codebooks) on eight test sets: LibriSpeech, Seed-TTS English and Mandarin, and FLEURS in German, French, Mandarin, Japanese and Korean.

  • Speed (encode / decode real-time factor, lower is faster, one M4 Max CPU thread, FP32):
    • LiquidCodec: 0.0028 / 0.0115
    • Mimi: 0.0361 / 0.0329
    • X-Codec 2: 0.7359 / 0.0379
    • NanoCodec: 0.2145 / 0.2543
  • Reconstruction: best ViSQOL and lowest mel distance on all eight sets, and it keeps that lead with a Q4_0 decoder. On LibriSpeech, ViSQOL is 4.319, vs 4.219 for NanoCodec, 4.119 for X-Codec 2 and 4.031 for Mimi.
  • TTS test: Liquid trained the same TTS model (LFM3-1.7B backbone, 101k hours, six languages, same batches) on each codec's tokens. LiquidCodec tokens gave the highest UTMOSv2 naturalness on all eight sets and the best WER/CER on four of eight.
    • LibriSpeech TTS WER: 3.15% vs 3.32% (NanoCodec), 4.19% (X-Codec 2) and 6.27% (Mimi).
    • Seed-TTS English TTS WER: 3.50% vs 3.95%, 5.65% and 6.15%.
    • Repeating the test with a Qwen3-1.7B backbone gave consistent results, per Liquid.

Caveats

  • Speech only: the model card says quality on music, singing and environmental sound hasn't been evaluated. Don't drop it into a music model and expect miracles.
  • Not a clean sweep: NanoCodec scores higher on PESQ and on speaker similarity (SECS) on most sets. X-Codec 2 and NanoCodec beat it on TTS error rate for Japanese, Korean and Mandarin FLEURS.
  • Codec speed isn't assistant latency: 87× real-time decoding is just the token-to-audio step. End-to-end response time still depends on the 2B backbone.
  • Benchmarks are Liquid's own, run on its hardware with its TTS recipe.
  • License: the LFM Open License v1.0 allows commercial use only for organizations under $10M annual revenue. Above that threshold, commercial use isn't licensed and you need a deal with Liquid. Read the LICENSE before you ship.
  • Native Transformers class: the card also shows a LiquidCodecModel import, but that class isn't in a released Transformers version yet (5.22.2 at publish time). Use the remote-code path above.

Conclusion

LiquidCodec is the open, 74M-parameter speech codec that sits under Liquid's LFM3 on-device voice models. It's small, streams in both directions, and is fast enough on one CPU thread that decoding barely touches your latency budget. Its parallel-predictable GFSQ tokens are the main reason to look at it if you train your own TTS or voice-agent model. If you just want local voice chat, run LFM3-Audio-2B, which already includes it. If you're building a speech model, LiquidCodec is a serious alternative to Mimi to benchmark this week. Check the $10M revenue line first.

—Anabel ♡