Liquid AI published LiquidCodec on November 12, a roughly 74M-parameter audio codec that sits underneath its new LFM3-Audio-2B and LFM3-Omni-2B models. Liquid AI, the lab behind the LFM family of small on-device models, says the codec decodes speech far faster than competing codecs while matching them on reconstruction quality. For anyone shipping a voice assistant that must run locally, the codec is the part that decides latency and memory.

A codec turns speech into discrete tokens a language model can predict, then turns predicted tokens back into sound. Liquid used Mimi, a third-party codec, in its first audio model. It swapped out only the detokenizer in LFM2.5-Audio-1.5B and got decoding 8 times faster. The tokenizer stayed borrowed, and Liquid says that capped token quality. So the lab built the whole thing itself.

The main design change is quantization. Mimi relies on residual vector quantization, where each codebook corrects the error left by the one before it, so predictions must happen in sequence. LiquidCodec uses grouped finite scalar quantization instead, which breaks that chain and lets a model predict several codebooks at once. The encoder and decoder are symmetric, and the encoder is fully causal, which suits streaming audio. The codec runs at 12.5 Hz on 24 kHz waveforms.

Training came in two stages. Liquid first used 264,000 hours of speech across 107 languages, with no adversarial losses, and the output carried “metallic” artifacts. It then fine-tuned on a narrower pool of 55,000 hours drawn from cleaner, full-band recordings, adding adversarial critics, and the artifacts faded. A final decoder-only pass with quantization-aware training shrinks the decoder by a factor of 3.5 in Q4_0 format, with little effect on audio quality.

The speed figures come from a single CPU thread at FP32 precision on an Apple M4 Max MacBook Pro. LiquidCodec encodes at about 361 times real time and decodes at about 87 times. Its decoder real-time factor is 0.0115, against 0.0329 for Mimi, 0.0379 for X-Codec 2, and 0.2543 for NanoCodec. On encoding, Mimi scores 0.0361 and LiquidCodec 0.0028, roughly a thirteenfold gap by our arithmetic.

Quality claims are broader but less clear-cut. Liquid reports the top ViSQOL score and the smallest mel distance across all eight test sets, which span LibriSpeech, Seed-TTS in English and Mandarin, and FLEURS in five more languages. In its text-to-speech comparison, trained on 101,000 hours covering six languages, models using LiquidCodec tokens had the top UTMOSv2 naturalness score on all eight sets. They had the best intelligibility on four of eight. NanoCodec beat it on PESQ and speaker similarity in several tables, so “matches” is the accurate word, not “beats.”

Every comparison in the post was run by Liquid, and the post cites no outside replication. Liquid did hold the training data, TTS architecture, and compute budget fixed across codecs, and it repeated the test with a Qwen3-1.7B backbone to show the gains do not depend on its own LFM3-1.7B model. Mimi was tested with 12 codebooks, a setting Liquid chose.

The release is usable today. LiquidCodec is on Hugging Face and loads through Transformers, and the LFM3-Audio-2B and LFM3-Omni-2B GGUF files come in three llama.cpp-ready builds: Q8_0, Q4_K_M, and Q4_0. Liquid’s pitch is that quicker speech generation with less memory lets developers build apps that sustain a conversation entirely on the device.

Teams with a local voice feature on the roadmap now have a codec to benchmark on their own hardware and languages. If the decode numbers hold on phone-class chips and not just an M4 Max, the cost of a voice turn on-device drops enough to change what fits in the product.

Liquid AI, “LiquidCodec: High-Fidelity Audio Tokenization Optimized for Speed,” Liquid AI Blog, November 12, 2026.