LFM3-Audio-2B-MLX-4bit
4-bit MLX version of LFM3-Audio-2B for Apple Silicon. The weights take 1.58 GB, plus 0.15 GB for the audio codec in audio_tokenizer/.
The language backbone uses 4-bit affine quantization with groups of 64, except for the token embedding (also the output layer) and the value and feed-forward down projections in the layers that llama.cpp's Q4_K_M upgrades, which use 6 bits. The audio encoder, audio projector, depth decoder and codebook embedding stay in BF16. The audio codec is the original one. It is the smallest variant and loses some accuracy, so 8-bit is the recommended trade-off.
Find architecture and usage details in the original model card.
Use with MLX-VLM
Requires MLX-VLM 0.7.11 or later:
pip install -U "mlx-vlm>=0.7.11"
# Speech recognition
python -m mlx_vlm.generate --model LiquidAI/LFM3-Audio-2B-MLX-4bit \
--system "Perform ASR in English." --prompt "" --audio <AUDIO PATH> \
--temperature 0 --max-tokens 2048
# Text to speech
python -m mlx_vlm.generate --model LiquidAI/LFM3-Audio-2B-MLX-4bit \
--system "Perform TTS in American English using a female voice." \
--prompt "Hello there!" --output-modality audio --output tts.wav
# Speech-to-speech chat
python -m mlx_vlm.generate --model LiquidAI/LFM3-Audio-2B-MLX-4bit \
--system "Respond with interleaved text and audio." --prompt "" \
--audio <AUDIO PATH> --output-modality audio --max-tokens 1024 --output answer.wav
Speech output samples with audio_temperature=1.0, audio_top_k=128 and audio_repetition_penalty=1.1, the defaults in generation_config.json. The original model card lists the voices, languages and other system prompts. More options in the MLX guide.