LFM3-Audio-2B-MLX-3bit
A 3-bit MLX quantization of Liquid AI's LFM3-Audio-2B, published by Kentucky Open Science for local inference on Apple Silicon with MLX Audio.
The language backbone uses affine 3-bit quantization with group size 64. Speech components retain their source precision, and the original speech codec is included in audio_tokenizer/. The model package is approximately 1.42 GB, including the tokenizer and codec.
Source and changes
- Original model: LiquidAI/LFM3-Audio-2B.
- Conversion source: LiquidAI/LFM3-Audio-2B-MLX-bf16, revision
452469a1afe66dd4932d0e9237603229fa606c0d. - Quantization: affine, 3 bits, group size 64, applied to the language backbone.
- Modified source files:
model.safetensors,model.safetensors.index.json, andconfig.json. The configuration records the quantization and removes the upstream remote-code mappings for the native MLX Audio runtime. Publication notices are included in these files' metadata and in NOTICE. - Other retained model assets, including the codec, tokenizer, generation settings, and license, match the pinned source revision.
This repository contains the model assets and a dependency lock file. The runtime implementation is provided by MLX Audio.
Run on Apple Silicon
The validated environment uses Python 3.12, MLX 0.32.4, and MLX Audio 0.6.5.
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install "mlx-audio==0.6.5" "mlx==0.32.4"
Generate speech from text:
python -m mlx_audio.sts.generate \
--model Kentucky-Open-Science/LFM3-Audio-2B-MLX-3bit \
--task tts \
--text "Hello from Kentucky Open Science." \
--gen-kwargs '{"language":"American English","voice":"female"}' \
--max-new-tokens 96 --seed 42 \
--output-path hello.wav
Respond to a recording with interleaved text and speech:
python -m mlx_audio.sts.generate \
--model Kentucky-Open-Science/LFM3-Audio-2B-MLX-3bit \
--task sts \
--audio question.wav \
--system-prompt "Respond with interleaved text and audio. Keep answers to one or two short sentences." \
--max-new-tokens 512 --seed 42 \
--output-path reply.wav
The runtime downloads the repository on first use. To download a portable local copy explicitly:
hf download Kentucky-Open-Science/LFM3-Audio-2B-MLX-3bit \
--local-dir ./LFM3-Audio-2B-MLX-3bit
Then replace the model identifier in the commands above with ./LFM3-Audio-2B-MLX-3bit. To reproduce the complete dependency environment, install the downloaded requirements.lock.txt in a Python 3.12 environment. WAV, MP3, and FLAC inputs are supported by the retained runtime; other formats may require ffmpeg. Speech output is mono at 24 kHz.
Validation and limitations
The release weights and index were checked for consistency. A short English text-to-speech smoke test generated a 24 kHz mono WAV with the pinned runtime. This verifies basic loading and speech generation; it is not an evaluation of speech quality, transcription accuracy, or all supported tasks and languages. No comparative quality or performance benchmark is included. Quantization may affect output quality.
The runtime supports speech recognition, text-to-speech, audio questions, and speech-to-speech chat. See the original model card for the base model's capabilities and limitations. Interleaved voice chat in the validated MLX Audio runtime is English only. This release targets Apple Silicon and MLX Audio; compatibility with other runtimes has not been validated.
License and attribution
The original model and weights are by Liquid AI. This quantized release is published by Kentucky Open Science. The weights retain the LFM Open License v1.0, including its redistribution conditions and commercial-use terms. Preserve the license and attribution notices when redistributing.