Liquid AI published a technical account of LiquidCodec on November 12, 2026, detailing the streaming speech tokenizer used in its on-device audio models. The disclosure explains its architecture and company-run codec comparisons, giving university speech researchers evidence to distinguish reconstruction fidelity from generated-speech quality.
The development follows the October 27 LFM3 family announcement, which already described LiquidCodec in LFM3-Audio-2B and LFM3-Omni-2B. November 12 marks the detailed engineering disclosure, rather than another launch of those models.
LiquidCodec uses a causal, symmetric architecture
According to its model card, LiquidCodec has approximately 74 million parameters and processes 24 kHz mono speech. Its fully causal encoder mirrors the decoder, supporting streaming encoding and decoding. Grouped finite scalar quantization divides features into independently quantized groups, allowing speech models to predict multiple codebooks in parallel. Each 80-millisecond frame contains 12 tokens.
The decoder predicts spectral coefficients directly, avoiding the learned progressive upsampling used in some other codecs. These choices target the computational constraints of local speech processing.
Controlled tests separate reconstruction from speech generation
In the November 12 disclosure, Liquid AI says it held training data, architecture and compute constant across text-to-speech experiments: 101,000 hours across six languages, identical batch order and three generation seeds. This tests how readily a speech model learns each codec’s tokens, separately from encoding and decoding an existing recording.
The company reports stronger ViSQOL and mel-distance reconstruction scores across eight test sets. For generated speech, it reports higher UTMOSv2 naturalness scores across all eight, while word or character error rates led on four. The results therefore support different conclusions about fidelity, naturalness and intelligibility; a reconstruction lead alone does not establish a generation lead.
The model card also reports decoding at about 87 times real time using FP32 inference on one Apple M4 Max CPU thread. That measures the codec stage, rather than a complete voice application’s response latency. For a reproducible AI experiment, the hardware, precision and measured stage are material parts of the result.
Weights are available under the LFM licence
The public codec repository contains BF16 weights for Transformers. Its documentation says quantized decoder variants are bundled into the LFM3-Audio-2B and LFM3-Omni-2B GGUF releases for llama.cpp, rather than distributed as separate codec checkpoints. The audio GGUF repository documents local CPU and GPU inference.
The repository’s LFM Open License v1.0 includes a commercial-use limitation. Section 5(c) exempts qualified nonprofit organisations’ noncommercial or research uses from its revenue threshold. For research offices and university technology-transfer teams, the distinction between research use and commercial deployment matters when assessing an open weights model.
The model card also states that quality on music, singing and environmental sound has not been evaluated. For universities, the disclosure supplies a speech-focused basis for comparison, with a defined evaluation scope rather than evidence covering every kind of audio research.
Sources
- Liquid AI: LiquidCodec technical disclosure, published 2026-11-12; checked 2026-11-14.
- Liquid AI: Introducing LFM3, published 2026-10-27; checked 2026-11-14.
- Liquid AI: LiquidCodec model card and weights; checked 2026-11-14.
- Liquid AI: LFM3-Audio-2B-GGUF model card and files; checked 2026-11-14.
- Liquid AI: LFM Open License v1.0; checked 2026-11-14.







