Not tested by us yet. Our test harness calls chat models over an API and cannot score audio, and a codec is not something you call as a service. Everything below comes from the LFM3-Audio-2B model card, Liquid AI's LFM3 announcement and launch coverage in AlphaSignal and MarkTechPost. We'll update this post when we can reproduce the speed numbers on our own hardware.
LiquidCodec is a streaming speech codec of about 74 million parameters from Liquid AI. It turns 24 kHz speech into 12 discrete tokens per 80 ms frame, and back. Its job is to be the audio tokenizer for voice models, and Liquid uses it inside LFM3-Audio-2B and LFM3-Omni-2B. If you build or fine-tune on-device voice models, it is worth a look. If you only call a hosted voice API, it changes nothing for you today.
| LiquidCodec | |
|---|---|
| Maker | Liquid AI |
| Reported | 2026-11-12 |
| Size | about 74M parameters |
| Audio | 24 kHz speech |
| Frame rate | 12.5 frames per second (80 ms per frame), 12 tokens per frame |
| Quantization | Grouped finite scalar quantization (GFSQ), not residual vector quantization |
| Encoder / decoder | Fully causal encoder built from LFM3 blocks; decoder predicts spectral coefficients directly |
| Access | Weights on Hugging Face, loadable through Transformers; bundled in the LFM3-Audio-2B repository |
| Price | None listed; open weights (the model card names the LFM Open License v1.0 for LFM3-Audio-2B) |
What a codec does, and why this one is different
A voice language model cannot read raw audio efficiently. A codec compresses speech into a short sequence of tokens the model can predict, then a decoder turns predicted tokens back into sound. Kyutai's Mimi did this job in Liquid's earlier LFM2-Audio-1.5B, per the coverage above. LiquidCodec replaces it in the LFM3 audio models.
Two design choices stand out in the reporting. GFSQ splits each feature vector into groups that are quantized independently, so a model can predict the tokens of one frame in parallel with MaskGIT-style refinement instead of one codebook at a time. And the encoder is causal, so it never needs future audio, which is what streaming needs. Decoding can start before the full reply is generated.
The numbers, and who reported them
Every figure below is Liquid's, measured by Liquid. We have not reproduced any of them.
| Claim | Value | Source |
|---|---|---|
| Encoding speed | about 361x real time, one CPU thread | AlphaSignal |
| Decoding speed | about 87x real time, one CPU thread | model card |
| Versus Mimi | about 13x faster encoding, about 3x faster decoding | AlphaSignal |
| Versus others | more than 250x faster encoding than X-Codec 2; 22x faster decoding than NanoCodec | AlphaSignal |
| Test hardware | one thread of an Apple M4 Max, FP32 | AlphaSignal |
| Quantized decoder | Q4_0 weights 3.5x smaller, with limited reported quality loss | AlphaSignal |
| ViSQOL on LibriSpeech | 4.319 versus 4.031 for Mimi | MarkTechPost |
The quality claim comes from a controlled comparison: the same 101,000 hours of training audio across six languages, the same LFM3-1.7B backbone, eight test sets drawn from LibriSpeech, Seed-TTS and FLEURS. Liquid reports the best ViSQOL and UTMOSv2 and the lowest mel distance on all eight sets, and the best word or character error rate on four of eight.
Read that last part closely. It means LiquidCodec did not win intelligibility on the other four. AlphaSignal notes that NanoCodec led several non-English intelligibility tests and scored better on some PESQ and speaker-similarity measures. The 361x and 87x figures also cover codec execution only. End-to-end latency still depends on buffering, token generation speed and playback.
What the speed does and doesn't mean
An encode at 361x real time means a codec step is unlikely to be your bottleneck on a laptop CPU. That matters for on-device assistants, where the language model and the codec share a small power budget. It does not tell you how a voice agent built on it feels to talk to. For that, the useful numbers are time to first audio and turn latency, and Liquid has not published those for the codec alone.
The practical catch is compatibility. LiquidCodec tokens are incompatible with Mimi, Encodec and SoundStream, so an existing speech model cannot swap it in. You would retrain or adapt the model. For most teams that makes LiquidCodec a component of Liquid's own models rather than a drop-in upgrade.
How the LFM3 audio models perform
The codec's parent models have more independent-looking numbers. On its model card, LFM3-Audio-2B lists a VoiceBench score of 73.13 (Gemma 4 E4B: 72.90) and an English text-to-speech UTMOSv2 of 3.83 (Qwen2.5-Omni 3B: 3.42). It trails on English speech recognition, with a word error rate of 6.23 against 5.51 for Voxtral Mini 3B, and on internal tool calling, 58.50 against 72.40 for Gemma 4 E4B. These are also vendor-run. Voice chat is English-only, though speech generation covers 18 languages.
If tool calling on small models matters more to you than voice, our BotPsy explainer covers a 0.8B function-calling model with the same caveats.
Who should care
- On-device voice builders: the combination of a causal encoder, a CPU-speed decoder and 4-bit decoder weights is aimed at you. Test it on your target hardware and your languages first.
- Researchers training speech LMs: GFSQ with parallel token prediction is the part to study. Check the license terms before building on it.
- Teams using hosted voice APIs: nothing to do yet. For a hosted multilingual option, see our look at Chariot-1-TTS-Multilingual.
We'll add a retest note once we can run LFM3-Audio-2B locally against a fixed set of scripts.