Liquid AI releases LiquidCodec: a 74M streaming speech codec that encodes at about 361x real time on a single CPU thread
Liquid AI has released LiquidCodec, a streaming speech codec of about 74 million parameters, with weights published on Hugging Face, intended to replace Mimi in LFM3-Audio-2B and LFM3-Omni-2B. The vendor self-reports encoding at about 361x real time on a single CPU thread; the report also credits it with the best ViSQOL and lowest mel distance across eight test sets.
What changed
According to a MarkTechPost report dated 12 November 2026, Liquid AI released LiquidCodec, a streaming speech codec of about 74 million parameters (the model card states 73.6M), with weights published on Hugging Face. It serves LFM3-Audio-2B and LFM3-Omni-2B and replaces Kyutai's Mimi — 96.2 million parameters, previously used in LFM2-Audio-1.5B. The report says the codec compresses 24 kHz audio into 12 token streams, with each 80-millisecond frame corresponding to 12 discrete tokens at a 12.5 Hz frame rate. The pipeline is encode, quantize, decode: the encoder is fully causal, which suits streaming; quantization uses grouped finite scalar quantization (GFSQ) rather than residual vector quantization (RVQ), so the 12 codebooks are independent of one another and can be predicted in parallel by a MaskGIT-style transformer; the decoder predicts spectral coefficients directly, skipping learned transposed-convolution upsampling. Training runs in two stages: the first uses 264,000 hours of speech in 107 languages with waveform and spectral losses only; the second fine-tunes on 55,000 hours of full-band speech with an adversarial discriminator, which Liquid AI says removes the metallic artifacts of the first stage; finally there is quantization-aware training (QAT) on the decoder only, aimed at edge inference. On speed, Liquid AI self-reports about 361x real time for encoding on one CPU thread of an Apple M4 Max laptop in FP32 (encoding RTF 0.0028, versus Mimi's 0.0361) and about 87x real time for decoding (decoding RTF 0.0115, versus Mimi's 0.0329). On quality, the report says LiquidCodec achieves the best ViSQOL and lowest mel distance across eight test sets, with ViSQOL 4.319 on LibriSpeech against NanoCodec 4.219, X-Codec 2 4.119 and Mimi 4.031; in a TTS comparison holding data, architecture and compute fixed, it has the highest UTMOSv2 across the eight sets, with LibriSpeech WER of 3.15% versus Mimi's 6.27%. For deployment, the model can be loaded through Hugging Face Transformers with trust_remote_code=True, calling encode() and decode() to obtain audio codes and waveform, and the native LiquidCodecStreamingCache supports chunk-by-chunk decoding; the model card example uses CUDA with bfloat16. On the edge side there are also Q4_0, Q4_K_M and Q8_0 GGUF versions for llama.cpp, released with the GGUF repositories for LFM3-Audio-2B and LFM3-Omni-2B; the report says Q4_0 with decoder-only QAT shrinks decoder footprint 3.5x while LibriSpeech ViSQOL drops only from 4.319 to 4.292.
What it means for developers
The most immediate effect is that the token layout changes. From the reported 12.5 Hz frame rate and 12 codebooks, one second of audio works out to roughly 150 discrete tokens (this is an inference from the reported figures, not the vendor's own wording). That means any TTS or speech language model trained on Mimi tokens cannot treat the codec as a hot-swappable component: vocabulary size, codebook count and the per-frame token sequence all differ, and checkpoints already fine-tuned on Mimi tokens need the corpus re-tokenized and the output head retrained, or quality will visibly degrade under the same backbone. GFSQ's independence changes the design space on the decoder side — the inter-codebook delay patterns common with RVQ are not needed here, so a speech language model can predict the 12 codebooks in parallel, and the number of decoding steps per frame may therefore fall substantially, at the cost of needing a head structure that predicts in parallel by group rather than autoregressing codebook by codebook. The CPU-side RTF figures affect service topology: the report says a single CPU thread encodes at about 361x real time and decodes at about 87x real time. If that order of magnitude roughly holds on your target hardware, the codec no longer has to occupy a GPU, which can be left to the language model, and the number of concurrent streams is set by CPU core count. But that figure comes from a single CPU thread in FP32 on an M4 Max laptop and is vendor-reported; moving to server CPUs and mixed real-world sampling rates requires re-measurement. On streaming latency, the encoder is described as fully causal while the source of the decoder's causality is not disclosed; the model card does provide a chunk-by-chunk decoding cache, but chunk size and first-audio latency still need to be measured, not assumed. The quality trade-off is not one-sided either: the report explicitly lists NanoCodec as higher on PESQ (WB) on seven of the eight test sets, and ahead on LibriSpeech speaker similarity with SECS 0.836 versus 0.785, so "best" holds only for specific metrics, and if the product goal leans toward preserving speaker identity the conclusion may reverse. Finally, weights being on Hugging Face does not mean a permissive license: the report says the model page is labelled other and the blog points to the LFM license, so terms should be checked before commercial use. LiquidCodec is itself a codec, not a chat model; it does not change how text or inference requests are integrated. What changes is the token contract, the training assets and the latency budget for the speech input/output segment.
Integration and evaluation
All of the following is suggestion and conditional analysis, not vendor statements, and needs verification on your own data and hardware. Decision one: whether to migrate existing speech assets. The reported change behind it is that LFM3-Audio-2B and LFM3-Omni-2B switch to LiquidCodec in place of Mimi. The choice to make: teams that have already trained models on Mimi tokens should schedule the codec swap as a tokenizer migration (re-tokenize and adapt or retrain the language model head), while entirely new projects can start from LiquidCodec tokens. The premise that would change this choice is whether you hold trained checkpoints or cached data bound to Mimi's vocabulary and codebook layout; if there are no reusable token-level assets, the cost drops to the evaluation level. A measurable acceptance criterion: on a fixed held-out audio set with the same backbone and the same training configuration, compare TTS WER/CER and UTMOSv2 before and after the migration, and separately measure codec round-trip ViSQOL once, confirming the drop falls inside your release threshold before proceeding. Decision two: whether the codec sits on CPU or GPU, and whether to quantize. The reported changes behind it are the self-reported codec speed on a single CPU thread, plus the llama.cpp-oriented Q4_0/Q4_K_M/Q8_0 GGUF builds and decoder-only QAT (decoder footprint 3.5x smaller, ViSQOL from 4.319 to 4.292). The choice to make is between putting the codec on CPU threads with a quantized decoder, or following the model card example down the CUDA bfloat16 path. The premise that would change this choice is concurrent stream count, the target device's memory ceiling, and whether you are more sensitive to first-audio latency or reconstruction quality; high-concurrency, memory-constrained edge deployments fit the former better, while the highest reconstruction quality and lowest single-stream latency fit the latter. Measurable acceptance criteria: end-to-end first-chunk/first-audio latency, concurrent streams each core can sustain, the decoder's actual memory footprint after quantization, and ViSQOL and mel distance on your own audio — if quantized metrics exceed your threshold, revert to a higher-precision tier. Decision three: which metric decides codec selection. The reported changes behind it are that LiquidCodec is best on ViSQOL and mel distance across all eight test sets and highest on TTS UTMOSv2, but trails NanoCodec 1 to 7 on PESQ (WB) and trails on speaker similarity with SECS 0.785 versus 0.836. The choice to make is to first determine whether the product cares more about content fidelity and intelligibility or speaker identity preservation, pick the primary metric accordingly, and keep NanoCodec as a comparison baseline where needed. The premise that would change this choice is whether the business scenario involves dubbing, cloning or speaker discrimination; pure transcription post-processing is insensitive to SECS, while voice-character products are highly sensitive. A measurable acceptance criterion is to measure SECS, PESQ (WB) and word error rate together on your own speaker set and judge each against its own threshold, rather than looking at a single aggregate score.
Limits and open questions
All speed and quality figures in the report come from Liquid AI's own testing; the speed was measured on a single CPU thread in FP32 on an Apple M4 Max laptop, and the eight test sets were chosen by the vendor too. None has been independently reproduced by a third party. The source does not disclose LiquidCodec's bitrate, nor whether the decoder is fully causal, so bandwidth planning and a lower bound on streaming decode latency cannot be derived from the report and can only be measured. There is one discrepancy in framing between the table and the benchmark: the feature table lists Mimi as 8 codebooks at 1.1 kbps, while the benchmark footnote says Mimi was measured with 12 codebooks, so comparability must be confirmed before comparing across tables. Weights being published on Hugging Face does not amount to a permissive license: according to the report, the model page is labelled other and the blog points to the LFM license, and the report does not elaborate on training data provenance or redistribution terms, which should be checked before commercial use or redistribution. LiquidCodec does not currently appear in the KimiSeek model catalog; this version is not listed on KimiSeek, and the weights and related repositories in the report belong to Liquid AI's release channel on Hugging Face. Availability is governed by the vendor, and this article makes no inference about its availability on this site. In addition, the report does not cover dedicated results for Chinese speech or for codec round-trip quality in Chinese-language scenarios; although the eight test sets include a Mandarin set, no separate figure is given, so teams building Chinese speech products should run their own tests.