gatilin/lfm3-omni-liquidcodec

LFM3 Omni LiquidCodec

Community packaging of the standalone audio_tokenizer/ from LiquidAI/LFM3-Omni-2B, revision 21d75123f86c1f7ffb4bef1ea202710c4e7ebe67. The safetensors file is an exact copy of the original source file, validated against its source LFS SHA-256. Original model and configuration code are included. LiquidCodec is a speech/audio codec, not a visual VAE, and does not contain the Omni language or visual model.

Loading

Install PyTorch, Transformers 5.21.1, accelerate and safetensors.

import sys, torch
from huggingface_hub import snapshot_download
root = snapshot_download('gatilin/lfm3-omni-liquidcodec')
sys.path.insert(0, root)
from component import load_model
codec = load_model(root)
# One channel, 24 kHz: replace with actual waveform in [-1, 1].
audio = torch.zeros(1, 4800)
with torch.inference_mode():
    encoded = codec.encode(audio).audio_codes
    reconstructed = codec.decode(encoded).audio_values

Encoding pads to a whole 1,920-sample frame. Consult validation.json for the tested conditions: a synthetic CPU encode/decode smoke test, not a measure of speech quality or full Omni model parity.

License

The source weights are distributed under the original LFM Open License 1.0, reproduced in LICENSE, including its commercial-use conditions. Source code files retain their own copyright and license notices.