Liquid AI Releases LiquidCodec: A 74M-Parameter Streaming Audio Codec for Real-Time Speech Models
Liquid AI has released LiquidCodec, the audio codec powering its LFM3-Audio-2B and LFM3-Omni-2B models. LiquidCodec is a ~74M-parameter streaming speech tokenizer that converts speech into tokens a language model can predict, then reconstructs audio from those tokens. Liquid AI built it around two goals: producing easy-to-predict tokens and enabling fast conversion back to speech. The weights are available on Hugging Face.
TL;DR
- Size: ~74M parameters (73.6M per the model card); 24 kHz audio, 12.5 Hz frames, 12 codebooks.
- Runs on: a single CPU thread (tested on an Apple M4 Max laptop in FP32); CUDA GPUs via Transformers; Q40, Q4KM, and Q80 GGUF builds for llama.cpp.
- Performance: best ViSQOL and lowest mel distance across all 8 test sets compared with X-Codec 2, NanoCodec, and Mimi.
- Best: ~361x real-time encoding on a single CPU thread (encode RTF 0.0028 vs. Mimi's 0.0361).
What Is LiquidCodec?
LiquidCodec is a Liquid AI speech codec that compresses 24 kHz audio into 12 token streams. Each 80 ms frame becomes 12 discrete tokens, according to the LFM3-Audio-2B model card. It replaces Kyutai's Mimi, which Liquid AI used in LFM2-Audio-1.5B. Mimi has 96.2M parameters.
How Does LiquidCodec Work?
[Original source content was truncated at this point. The remaining sections on architecture, benchmark methodology, and deployment details could not be recovered from the provided text.]
2026 Context
LiquidCodec arrives as speech-native language models increasingly rely on compact, low-latency tokenizers to operate in real time on edge and CPU-only hardware. With single-thread CPU encoding roughly 361x faster than real time and GGUF builds available for llama.cpp, the codec reflects 2026's broader shift toward lightweight, quantized audio pipelines that can be deployed without dedicated accelerators.
via MarkTechPost