Liquid AI has released LiquidCodec, its proprietary 74M-parameter streaming audio codec, which serves as the foundation for its LFM3 audio models. The codec converts speech into discrete tokens and back faster than real-time directly on a standard laptop processor, without requiring a GPU or cloud infrastructure. The codec weights and quantized decoders are available on Hugging Face under the LFM Open v1.0 license.



What happened
On November 12, 2026, Liquid AI published the LiquidAI/Liquid-Codec repository on Hugging Face: the codec is distributed under the LFM Open v1.0 license, with weights provided in BF16 format. LiquidCodec operates at 24 kHz and outputs 12.5 frames per second with 12 discrete tokens, resulting in a bitrate of approximately 1.65 kbps. Instead of the residual vector quantization (RVQ) used in the Mimi codec, it employs Grouped Finite Scalar Quantization (GFSQ), allowing the language model to predict all 12 codebooks of a frame in parallel. On a single CPU thread of an Apple M4 Max, encoding runs 361 times faster than real-time (RTF 0.0028), while decoding runs 87 times faster (RTF 0.0115), compared to 0.0329 for Mimi, 0.0379 for X-Codec 2, and 0.2543 for NanoCodec. In terms of reconstruction quality, the codec achieved a ViSQOL of 4.319 on LibriSpeech and led in ViSQOL and mel-spectrogram distance across all eight test sets recorded in English, Chinese, German, French, Japanese, and Korean. The 4-bit Q4_0 decoder is 3.5 times smaller than the full version, with a quality loss of approximately 0.6%: ViSQOL drops from 4.319 to 4.292.
Context
In modern audio-LLMs, the audio codec acts as a bridge between continuous sound and the language model: it compresses speech into discrete tokens that the model manipulates like words, and expands generated tokens back into sound. The classic residual vector quantization scheme, used by the previous LFM2-Audio model on top of the Mimi codec, creates sequential dependencies: codebooks within a frame are predicted one after another, and each additional autoregressive step slows down speech generation. Abandoning sequential prediction reduces the number of steps during training and inference, thereby impacting the entire architecture of speech models, not just codec speed. The LiquidCodec architecture is designed for streaming scenarios from the outset: the encoder is fully causal and only sees the past, while the decoder contains no learnable upsampling and immediately outputs spectral coefficients. The LFM3 family is based on a foundation of 264,000 hours of speech in 107 languages, laying the groundwork for subsequent speech models, including localized and low-resource language versions. The codec release relies on the familiar open-source path of the llama.cpp ecosystem, where the community ports models into quantized formats for CPU execution.
Why this matters for the industry
For the industry, this event confirms that a proprietary audio codec has become an engineering lever rather than a marketing point: the speed of tokenization and detokenization determines whether voice input and output can be maintained on-device, and here these operations are moved from the "requires GPU or cloud" category to the budget of a standard laptop. Teams building on-device pipelines for speech recognition and synthesis receive an open building block without the cost of GPU farms, lowering the barrier to entry for voice products for startups and small teams. Cloud voice API providers face direct competition from open local solutions. If independent measurements confirm the stated metrics, GFSQ tokenization will likely become the default candidate for new audio-LLMs; community-retrained variants of the codec for other frequencies and languages, native support in edge stacks, and integrations into voice agent frameworks are expected. In the strategic perspective, the combination of "proprietary codec plus compact LLM" may become the standard recipe for on-device multimodality, where competition shifts from model size to the speed of processing audio tokens on CPU.
Why this matters for users
Developers and experienced readers have access to everything today: LiquidCodec BF16 weights can be downloaded from Hugging Face and run through Transformers with the trust_remote_code=True parameter, while quantized decoders Q4_0, Q4_K_M, and Q8_0 in GGUF format are located in the LFM3-Audio-2B and LFM3-Omni-2B repositories and can be run via llama.cpp on a laptop or phone. Based on this, a local prototype of a voice assistant can be assembled, where speech synthesis and recognition occur on the device, eliminating the per-call cost of cloud APIs; such prototypes can be built in days, not months. Open comparisons with Mimi, X-Codec 2, and NanoCodec allow users to re-verify vendor numbers on their own recordings and determine how much more noticeable the new codec is compared to competitors in real-world scenarios. If measurements are confirmed, the familiar pattern of "instant local voice input and output" will no longer require a separate graphics card.
What is currently unknown / limitations
All speed and quality figures are vendor-provided: they are cited in the Liquid AI blog and model cards, taken without a disclosed methodology on company equipment, so results may vary significantly on other processors and configurations. As of the publication of this material, there have been no public independent reproductions of the stated RTF and ViSQOL on third-party data, including noisy speech, telephone channels, and accents. Quality checks were conducted on six languages, and Russian is not among them, so the codec's performance on Russian speech remains an open question. Assessments regarding the transfer of GFSQ to future audio-LLMs and the displacement of cloud voice APIs from mass scenarios are interpretations, not company promises.
Sources
- LiquidCodec: High-Fidelity Audio Tokenization Optimized for Speed (Liquid AI Blog)
- LiquidAI/Liquid-Codec (Hugging Face model card)
- LiquidAI/LFM3-Audio-2B (Hugging Face model card)
Author
Look at AI, editorial team
