Liquid AI

Build efficient general-purpose AI at every scale.

[Liquid AI @liquidai]

Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.

> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒

Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.

1/ 🧵

[JJ @JosephJacks_]

Incredible ship by @liquidai 🧏🏻‍♀️ 🌊

[Viviana Márquez @vivmarquez]

Day 1 at the @sgl_project summit!

If you weren’t there today, you missed out on @tugot17’s talk, but you can still come say hi at our @liquidai booth tomorrow!

[Alexander Amini @xanamini]

LLM next-token prediction scales because text comes from a discrete, fixed vocabulary.

Audio doesn't - it's a continuous waveform signal. @liquidai LiquidCodec is our new tokenizer for sound, unlocking audio generation directly within discrete autoregressive language models.

Convert continuous audio into a discrete vocabulary, train on the tokens, and decode back the raw audio.

waveform → tokens → waveform 🚀

[▲ Marktechpost AI @Marktechpost]

▲ Liquid AI just released LiquidCodec, a ~74M-parameter streaming audio codec that turns speech into tokens a language model can predict, then back into audio. It powers LFM3-Audio-2B and LFM3-Omni-2B.

The key idea is grouped finite scalar quantization (GFSQ) instead of residual vector quantization. Each 80 ms frame becomes 12 independent tokens, so all 12 can be predicted in parallel. The decoder predicts spectral coefficients directly, skipping transposed-convolution upsampling.

On 1 CPU thread of an Apple M4 Max (FP32), it encodes at ~361x and decodes at ~87x real time. Encode RTF is 0.0028 vs 0.0361 for Kyutai's Mimi. It posts the best ViSQOL and TTS UTMOSv2 on all 8 test sets. It is a trade-off though: NVIDIA's NanoCodec still scores higher PESQ (WB) on 7 of 8 sets (LibriSpeech 2.942 vs 2.640), and Liquid AI has not published a bitrate or latency in ms.

Weights are on Hugging Face, with Q4_0, Q4_K_M and Q8_0 GGUF builds for llama.cpp. Q4_0 shrinks the decoder 3.5x......

Full analysis: https://t.co/TLJEOMRDCA

Model: https://t.co/HaR8eh06cM

Blog: https://t.co/Ff5uHr1ean

@liquidai

[Richie Suarez @liquid_richie]

LiquidCodec is out. It's the ~74M param audio tokenizer under LFM3-Audio-2B and LFM3-Omni-2B.

The part I like as an engineer: grouped FSQ instead of RVQ, so there's no chain between codebooks and a speech model can predict all 12 tokens of a frame in parallel.

One FP32 CPU thread on an M4 Max: ~361x real time encode, ~87x decode. Best ViSQOL and lowest mel distance on all 8 test sets, and the lead holds with a Q4_0 decoder.

Nice work Anna Piunova and Calder Katyal.

https://t.co/r3Jfj5ixzM

[Anya @anyapiunova]

And here it goes! LiquidCodec is out!

Built for high-quality speech generation and fast, efficient inference on edge devices, just like the rest of the Liquid family 🚀

[AI.Card.Daily @ai_card_daily]

Liquid AI가 자체 음성 코덱 LiquidCodec을 공개했다. 파라미터 약 7,400만 개의 작은 모델로, 10월 말 나온 LFM3-Audio-2B와 LFM3-Omni-2B에 들어간 음성 토크나이저다.

코덱은 말소리를 언어 모델이 예측할 수 있는 토큰으로 바꾸고, 모델이 만든 토큰을 다시 소리로 되돌린다. Liquid AI는 토큰이 예측하기 쉬울 것, 소리로 되돌리는 속도가 빠를 것, 이 두 조건을 두고 만들었다.

이전 오디오 모델(LFM2-Audio-1.5B)은 외부 코덱 Mimi를 썼고, 후속 LFM2.5-Audio-1.5B도 디토크나이저만 바꾸고 Mimi 토크나이저는 그대로 썼다. Mimi가 쓰는 RVQ 방식은 각 코드북이 앞 코드북이 남긴 오차를 차례로 메우는 구조라, 코드북끼리 서로 강하게 얽혀 있다.

LiquidCodec의 핵심은 세 가지다.

1. 서로 기대지 않는 토큰. 특징을 여러 그룹으로 나눠 따로 토큰으로 바꾸는 방식(GFSQ)을 써서, 코드북끼리 순서에 매이지 않는다. 그래서 음성 모델이 한 프레임의 토큰 12개를 동시에 예측할 수 있다.

2. 대칭 구조. 인코더와 디코더를 같은 모양으로 만들고 인코더도 앞부분만 보고 처리하게 해, 실시간 스트리밍에 맞췄다.

3. 두 단계 학습. 107개 언어 26만 4천 시간 분량으로 먼저 배우고, 깨끗한 음성 5만 5천 시간으로 다듬어 금속성 잡음을 줄였다.

Liquid AI에 따르면 속도는 Apple M4 Max CPU 스레드 1개, FP32 기준으로 실시간 대비 인코딩 약 361배, 디코딩 약 87배다. 모델 카드에 따르면 X-Codec 2, NanoCodec, Mimi와 비교한 8개 복원 테스트셋 모두에서 ViSQOL 점수가 가장 높고 mel 거리가 가장 낮았다. 다른 지표에서는 비슷한 수준이었다. 같은 조건의 음성 합성 실험에서도 LiquidCodec 토큰으로 학습한 모델이 8개 테스트셋 모두에서 자연스러움(UTMOSv2) 점수가 가장 높았다. 발음 정확도(WER/CER)는 8개 중 4개에서 가장 좋았다.

Liquid AI는 디코더만 양자화 인식 학습(QAT)을 거쳐 4비트(Q4_0)로 줄여도 품질 손실이 적고 용량은 3.5배 작아진다고 밝혔다. 다만 음악이나 노래, 환경음에서는 아직 평가하지 않았다. BF16 가중치는 허깅페이스에 공개됐고, 4비트 디코더는 LFM3-Audio-2B·LFM3-Omni-2B GGUF 안에 들어 있다.

#LiquidAI #LFM3 #HuggingFace

https://t.co/KHNaZTn5oH

[Alano @Alanogrok]

Reconstruction fidelity and learnable speech tokens are different objectives. Holding the TTS setup fixed makes that comparison more useful; streaming response latency still needs measuring beyond codec throughput.

[Nathan Ranchin @nathanrchn]

“blazingly fast audio codec” -> ⚡️🔉🤖

[Saniya Karwa @saniwhya]

this is what gave omni it's voice :) thanks to @anyapiunova & @CalderKatyal for leading this

[Jimmy Smith @jimmysmith1919]

We are releasing LiquidCodec, our in-house audio codec used by our LFM3 Audio and Omni models.

[Mathias Lechner @mlech26l]

Convert speech into tokens and back again

[leonie @helloiamleonie]

What is an 𝗮𝘂𝗱𝗶𝗼 𝗰𝗼𝗱𝗲𝗰?

(And why our audio team built their own)

𝗪𝗵𝗮𝘁 𝗶𝘀 𝗮𝗻 𝗮𝘂𝗱𝗶𝗼 𝗰𝗼𝗱𝗲𝗰?
1. Turns speech into tokens
2. Turns the tokens back into audio

𝗛𝗼𝘄 𝗱𝗼𝗲𝘀 𝗮𝗻 𝗮𝘂𝗱𝗶𝗼 𝗰𝗼𝗱𝗲𝗰 𝘄𝗼𝗿𝗸?
→ Encode: turn raw audio into feature vectors
→ Quantize: map those vectors into discrete tokens
→ Decode: reconstruct the waveform from the tokens

𝗪𝗵𝘆 𝗯𝘂𝗶𝗹𝗱 𝗼𝘂𝗿 𝗼𝘄𝗻?
Our earlier audio models used a third-party tokenizer.
But over time, its token quality and design choices started to limit us.

𝗟𝗶𝗾𝘂𝗶𝗱𝗖𝗼𝗱𝗲𝗰 is a ~74M-parameter streaming codec with:
• 𝗛𝗶𝗴𝗵-𝗳𝗶𝗱𝗲𝗹𝗶𝘁𝘆: reconstructed audio is closer to the original.
• 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗲𝗱 𝗳𝗼𝗿 𝘀𝗽𝗲𝗲𝗱: It's fast, even on-device.

Learn more about how our audio team built it in our newest engineering blog:
https://t.co/lL3wYzB01Q

[Maxime Labonne @maximelabonne]

it's a good codec, sir

[Vishnu Samavedula @svrvishnu]

We built our own codec.

[Ramin @ramin_m_h]

a 74M parameter streaming audio codec. You can enable many applications with it:

1) offline voice assistants for cars, appliances and phones;

2) phone support agents that handle interruptions gracefully, since the encoder never waits on future audio

3) live speech translation in earbuds and on calls, backed by training data in 107 languages

4) multilingual narration and TTS for audiobooks, e-learning and IVR

5) low-bandwidth voice for satellite messengers, push-to-talk radios and IoT intercoms that send tokens instead of waveforms

6) compact call and voicemail archives, with an hour of audio encoding in about 10 seconds on a single M4 Max CPU thread by Liquid's benchmark

7) cheap tokenization of huge speech corpora to train your own TTS or speech LM

8) game and XR characters that speak from the player's own device

9) offline assistive speech like voice banking and AAC devices, helped by a 4-bit decoder build that is 3.5× smaller

10) call-quality repair, where a small model predicts the tokens lost to dropped packets instead of playing silence.

Weights on HF, more: https://t.co/42IL0NI9Xn

[Liquid AI @liquidai]

High-quality voice interaction on a device needs fast audio processing and a small memory footprint.

Building LiquidCodec in-house lets us improve the speech representation alongside the models that learn from it.

LiquidCodec is available on Hugging Face. See the model card for installation and loading instructions.

Explore the architecture, training recipe, and results.

> Blog: https://t.co/yUf5TY74Fl
> Hugging Face: https://t.co/tkuoB6dLAO

[Liquid AI @liquidai]

Introducing 𝙻𝚒𝚚𝚞𝚒𝚍𝙲𝚘𝚍𝚎𝚌, our ~74M-parameter streaming audio codec powering LFM3-Audio-2B and LFM3-Omni-2B.

An audio codec turns speech into tokens a language model can predict, then converts those tokens back into audio. We built ours around two requirements: tokens that are easy to predict and fast conversion back into speech.

On a single CPU thread on an Apple M4 Max, LiquidCodec runs at approximately:
> 361× real time for encoding
> 87× real time for decoding

Both measurements use FP32 precision.

1/ 🧵

[Liquid AI @liquidai]

LiquidCodec’s speed comes from its architecture and how it reconstructs audio.

The encoder and decoder use the same LFM3 block architecture, and the encoder is fully causal for streaming. The decoder predicts spectral coefficients directly, avoiding the progressive upsampling with learned transposed convolutions used by some other codecs.

We also fine-tuned the decoder with quantization-aware training for on-device inference.

Q4_0 reduces the decoder footprint by 3.5× with minimal impact on reconstruction quality. It retains its ViSQOL and mel-distance lead over the external codecs in our comparison.

5/