LFM3-Audio-2B — CMF

Liquid AI LFM3-Audio-2B — a 2B speech model on LFM3: a causal audio encoder (16 kHz) feeding the LFM3-1.7B text model, which answers in text or in 24 kHz speech through a depth decoder and the LiquidCodec tokenizer (12 codes per 80 ms) — speech recognition, speech synthesis in 19 languages, voice chat and audio understanding, packaged as single CMF files with the codec inside. They run with cortiq, a Rust inference engine with no Python and no ML framework: NVIDIA, AMD and Intel GPUs through Vulkan/DX12, Apple silicon through Metal, and the CPU.

English · Русский · 中文

Quick start

cargo install cortiq-cli
hf download infosave/LFM3-Audio-2B-cmf lfm3-audio-2b-q8_2f.cmf --local-dir .

# speech recognition
cortiq run lfm3-audio-2b-q8_2f.cmf --system "Perform ASR in English." --audio clip.wav --prompt ""

# speech synthesis (24 kHz WAV)
cortiq run lfm3-audio-2b-q8_2f.cmf --system "Perform TTS in American English using a female voice." \
  --prompt "The quick brown fox jumps over the lazy dog." --speech-out tts.wav --max-tokens 512

Use cortiq 0.8.50 or later; prebuilt binaries for Linux, macOS and Windows are on the releases page. --audio takes a WAV file (other formats through ffmpeg), a data URL or an HTTP(S) URL and may be repeated; clips are resampled to 16 kHz mono. --system sets the task, as in Liquid's examples. --speech-out writes the spoken answer; --max-tokens counts its frames (12.5 a second).

Files

Against the bf16 checkpoint run in float32 by Transformers:

• Speech input. The log-mel features, the encoder rows (cosine 1.000000 for every row) and the rows handed to the language model (largest difference 4e-6) match; both files transcribe the test clips word for word as Transformers does.
• Speech output. With greedy codes in a fixed codebook order, an f16 conversion reproduces Transformers' speech frames exactly (42 of 42 frames, every code). The quantized files take other greedy codes after the first frames — a near tie over 2,017 codes — and their speech transcribes back to the input sentence word for word.

Performance

cortiq run, cortiq 0.8.50, q4tp:

A speech frame is one step of the language model (on the GPU) plus three passes of the depth decoder (on the CPU, 16–23 ms a frame); the codec decodes the frames at the end (0.1 s on the M4, 0.6 s on the A40 host's CPU for 7 s of speech).

Usage

# voice chat: interleaved text and speech
cortiq run lfm3-audio-2b-q8_2f.cmf --system "Respond with interleaved text and audio." \
  --audio question.wav --prompt "" --speech-out reply.wav --speech-mode mixed --max-tokens 1024

# questions about a recording
cortiq run lfm3-audio-2b-q8_2f.cmf --audio street.wav --prompt "What can be heard in this clip?" --no-think

# speech in another language and voice
cortiq run lfm3-audio-2b-q8_2f.cmf --system "Perform TTS in German using a male voice." \
  --prompt "Guten Morgen! Wie geht es dir?" --speech-out de.wav

TTS languages: Arabic, Danish, German, Greek, British English, American English, Finnish, French, Hebrew, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Spanish, Swedish, Turkish, Chinese; female or male voice. Speech is sampled as Liquid recommends (temperature 1.0, top-k 128, repetition penalty 1.1 over 4096 frames); --greedy takes the likeliest codes. Text answers never open a speech turn.

OpenAI-compatible server

cortiq serve lfm3-audio-2b-q8_2f.cmf --port 8080

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "lfm3-audio", "messages": [
         {"role": "system", "content": "Perform ASR in English."},
         {"role": "user", "content": [{"type": "input_audio",
           "input_audio": {"data": "<base64 WAV>", "format": "wav"}}]}]}'

Audio parts take input_audio (base64), an HTTP(S) or data: URL, or a local path; the server answers in text. Speech output (--speech-out) is a run feature for now, as are live (streaming) recognition and speech in the server.

Hardware

NVIDIA / AMD / Intel (Vulkan, DX12). Both files fit any card with 3 GB of free memory; on headless Linux install libvulkan1 libglvnd0 libegl1 libgl1 libglx0 and set XDG_RUNTIME_DIR=/tmp. Apple silicon: the Metal backend is the default. CPU only: CMF_GPU=0.

Verify the download

sha256sum -c lfm3-audio-2b-q8_2f.cmf.sha256
cortiq verify lfm3-audio-2b-q8_2f.cmf

Reproduce

cortiq convert --model LiquidAI/LFM3-Audio-2B --quant q8_2f \
  --output lfm3-audio-2b-q8_2f.cmf --also-output q4tp=lfm3-audio-2b-q4tp.cmf

The converter reads the codec from the checkpoint's audio_tokenizer/ folder into the same file. Scripts (Transformers reference, speech parity): scripts/lfm3.

License

These files are a modified form of Liquid AI's LFM3-Audio-2B: the weights were converted to CMF and the language model quantized (8-bit q8_2f, 4-bit q4tp) by cortiq. They remain under the LFM Open License v1.0, a copy of which is in this repository (LICENSE); among its terms, commercial use by an entity with annual revenue of US$10 million or more is not licensed. See the upstream card.

---

Документация на русском

LFM3-Audio-2B от Liquid AI — речевая модель на 2 млрд параметров поверх LFM3: причинный аудиоэнкодер (16 кГц) и текстовая LFM3-1.7B, отвечающая текстом или речью 24 кГц через depth-декодер и кодек LiquidCodec (12 кодов на 80 мс): распознавание речи, синтез на 19 языках, голосовой чат и понимание аудио. Файлы CMF включают кодек и запускаются движком cortiq на Rust без Python.

cargo install cortiq-cli
hf download infosave/LFM3-Audio-2B-cmf lfm3-audio-2b-q8_2f.cmf --local-dir .
cortiq run lfm3-audio-2b-q8_2f.cmf --system "Perform ASR in Russian." --audio clip.wav --prompt ""
cortiq run lfm3-audio-2b-q8_2f.cmf --system "Perform TTS in American English using a female voice." \
  --prompt "Hello! How are you?" --speech-out tts.wav

Нужен cortiq 0.8.50 или новее. Распознавание совпадает с Transformers дословно (строки энкодера — косинус 1.000000); синтез в f16 воспроизводит кадры речи Transformers точно, речь квантованных файлов распознаётся обратно в исходную фразу. 7 с речи синтезируются за 5.6 с на NVIDIA A40 и за 3.2 с на Apple M4. --speech-mode mixed — голосовой чат (текст и речь вперемешку); cortiq serve принимает аудио в формате OpenAI (input_audio) и отвечает текстом. Файлы — изменённая форма LFM3-Audio-2B от Liquid AI по LFM Open License v1.0 (копия — LICENSE); коммерческое использование организацией с годовой выручкой от 10 млн долларов США лицензией не покрывается.

中文说明

LFM3-Audio-2B 是 Liquid AI 基于 LFM3 的 2B 语音模型：因果音频编码器（16 kHz）接 LFM3-1.7B 文本模型，可用文本或 通过 depth 解码器与 LiquidCodec 编解码器（每 80 ms 12 个码）输出 24 kHz 语音—— 语音识别、19 种语言的语音合成、语音对话与音频理解。本仓库的 CMF 文件内含编解码器， 可由 Rust 推理引擎 cortiq（0.8.50 及以上）加载。

cargo install cortiq-cli
hf download infosave/LFM3-Audio-2B-cmf lfm3-audio-2b-q8_2f.cmf --local-dir .
cortiq run lfm3-audio-2b-q8_2f.cmf --system "Perform ASR in Chinese." --audio clip.wav --prompt ""
cortiq run lfm3-audio-2b-q8_2f.cmf --system "Perform TTS in Chinese using a female voice." \
  --prompt "你好，今天天气很好。" --speech-out tts.wav

识别结果与 Transformers 逐字一致；f16 转换的语音帧与 Transformers 完全相同，量化文件 合成的语音可被逐字识别回原句。7 秒语音在 NVIDIA A40 上 5.6 s、在 Apple M4 上 3.2 s 合成完毕。这些文件是 Liquid AI LFM3-Audio-2B 的修改形式，仍受 LFM Open License v1.0 约束（副本见 LICENSE）；年收入 1000 万美元及以上的实体进行商业使用 不在许可范围内。