Liquid AI Open-Sources 74M-Parameter Streaming Audio Codec LiquidCodec

Liquid AI releases 74M streaming audio codec LiquidCodec, ~361x real-time on a single CPU thread.
Liquid AI has released LiquidCodec, a streaming audio codec with ~74M parameters, serving as the underlying component for LFM3-Audio-2B and LFM3-Omni-2B, responsible for discrete encoding and reconstruction of audio. It balances compression rate and audio quality with a 12.5 Hz frame rate and 12 codebooks, and achieves encoding speeds of ~361x real-time on a single CPU thread, suitable for large-scale preprocessing and edge deployment. The article notes that codecs determine how audio is tokenized and understood by language models, and both models sharing the same scheme implies a unified audio input/output interface. The streaming design targets low-latency scenarios such as real-time dialogue and transcription, but subjective audio quality, horizontal comparisons, and latency performance under network jitter have not yet been disclosed.
Liquid AI has released a streaming audio codec called LiquidCodec with approximately 74M parameters. This model is not a standalone product for end users, but rather an underlying component that provides audio encoding and decoding capabilities for the LFM3-Audio-2B and LFM3-Omni-2B audio models.

Design Trade-offs: Small Size and High Throughput
LiquidCodec's core specifications center on "real-time streaming processing." It operates at a frame rate of 12.5 Hz and uses 12 codebooks to represent audio information. This combination means the model only needs to process a small number of time steps per second, while preserving sufficient acoustic detail through parallel multi-codebook representation, achieving a balance between compression rate and reconstruction quality.
More noteworthy is the efficiency performance. According to official figures, LiquidCodec can achieve encoding speeds of approximately 361x real-time on a single CPU thread. What this number means: processing one minute of audio requires less than one second of computation time, without relying on GPU acceleration. For scenarios requiring large-scale audio preprocessing or deploying voice interaction on edge devices, this CPU-friendly characteristic has practical value.
Basic Meaning of Codebooks and Frame Rate
In neural audio codecs, a "codebook" refers to a discrete vector table. The encoder maps each frame of audio to several codebook indices, and the decoder reconstructs the waveform accordingly. A single codebook has limited representational capacity, so parallel multi-codebook approaches (such as residual vector quantization or grouped quantization) are commonly used, employing multiple indices together to characterize the acoustic details of the same frame—the more codebooks, the richer the audio quality and timbre information that can be preserved, but the longer the token sequence.
"Frame rate" refers to the number of time steps segmented and encoded per second. The lower the frame rate, the shorter the sequence, and the fewer tokens the language model or downstream network needs to process, reducing computation and memory overhead; the cost is that each frame must carry more information, typically relying on more codebooks for compensation. 12.5 Hz means one time step every 80 milliseconds, a typical low-frame-rate streaming configuration, which paired with 12 codebooks forms a "short sequence, wide representation" trade-off.
Why Codecs Become a Key Link in Audio Models
In the context of large models, codecs are often less attention-grabbing than the models themselves, but they determine how audio signals are discretized and how they are "understood" by language models. LiquidCodec serves both LFM3-Audio-2B and LFM3-Omni-2B, indicating that Liquid AI is unifying audio input and output onto a shared tokenization scheme.
The combination of multi-codebook design and low frame rate is a common approach in current streaming audio generation: low frame rate reduces sequence length and lightens the model's computational burden; multiple codebooks compensate for the information loss caused by low frame rate. 12 codebooks is a relatively rich configuration, typically pointing to applications with higher audio quality requirements, rather than mere speech recognition preprocessing.
Discrete Tokenization and Integration with Language Models
Current audio large models generally borrow the modeling paradigm of language models: first quantizing continuous waveforms into discrete token sequences, then modeling them with autoregressive or masked prediction just like processing text. The codec thus plays the role of a "tokenizer," and its codebook structure directly determines the vocabulary form and sequence length of tokens. If multiple models share the same codec, audio input and output have a unified token space, facilitating speech understanding, generation, and full-duplex dialogue within the same framework, and also saving the cost of training a separate tokenizer for each model. LiquidCodec serving both LFM3-Audio-2B and LFM3-Omni-2B is precisely a manifestation of this shared tokenization scheme.
Streaming Capabilities Point to Real-Time Interaction Scenarios
"Streaming" is the keyword for this codec. Real-time voice dialogue, live captioning, online meeting transcription, and other scenarios all require audio processing to be conducted in a low-latency, continuous manner, rather than waiting for an entire audio segment to end before batch processing. LiquidCodec's 12.5 Hz frame rate naturally adapts to chunked processing, and combined with high throughput on a single CPU thread, it can maintain a stable real-time processing pipeline in resource-constrained environments.
Currently, public information is concentrated at the specification level, with no disclosure yet of subjective audio quality scores, horizontal comparisons with other codecs, or latency performance under real network jitter conditions. These details will determine its effectiveness in actual product deployment.
The Difference Between "Real-Time Multiple" and Streaming Latency
361x real-time (the reciprocal of 361× RTF) measures throughput: the computation time required to process a unit duration of audio, reflecting computational efficiency, and is not equivalent to interaction latency. The true streaming experience also depends on whether the algorithm needs to see future context, the chunk window size, and the additional delay introduced by the encoding-decoding round trip. Low-frame-rate chunking facilitates computing while receiving, but if the model relies on cross-chunk context, end-to-end latency may still accumulate. Therefore, when evaluating such codecs, throughput data and first-packet latency and jitter robustness need to be considered separately.
Summary
The release of LiquidCodec completes the underlying layer of Liquid AI's audio model stack. The 74M parameter count, 12.5 Hz frame rate, 12 codebooks, and encoding speed of approximately 361x real-time on a single CPU thread together outline a lightweight audio codec solution for real-time deployment. Its actual impact will ultimately be tested by the performance of LFM3-Audio-2B and LFM3-Omni-2B in voice interaction tasks.
Related articles

Abnormal Error Rate Spike in Claude Mythos 5.2 and Fable 5.2
Claude Mythos 5.2 and Claude Fable 5.2 show an abnormal spike in request error rates, marked as investigating on the official status page. This article covers the incident overview, practical impact on developers, and recommended responses.

ReAct Runtime Context Trimming: Why Claude Code Acts at 70%
A deep dive into ReAct runtime context trimming: why Claude Code compresses at ~70% usage, why tool calls dominate, why messages must be dropped in pairs, and why codebases don't need RAG.

AI Agent vs LLM: How Cursor and Claude Code Actually Work
An in-depth analysis of the essential difference between AI Agents and LLMs: LLMs generate code suggestions, while Agents modify files and run tests via tool calls. Using a Spring Boot bug fix as an example, it breaks down the workflows, permission control, and verification mechanisms of Cursor and Claude Code.