nativ-community/LFM3-Audio-2B-4bit

LFM3-Audio-2B · 4bit

Native MLX conversion of lq-preview/LFM3-Audio-2B at 74fafd3bf0aaba3a5927a71cdc543c38d794bb6f. Conversion implementation: 98c05bd33cc93b64b12bb19e688a1511bbd4f82c (recorded in conversion.json). Public inference code: Blaizzy/mlx-vlm.

Weights: affine 4-bit, group size 64. Audio and vision modules remain BF16; the bundled audio codec is copied unchanged. Activation and KV caches are not quantized.

Usage

Install the pinned public implementation:

pip install 'mlx-vlm @ git+https://github.com/Blaizzy/mlx-vlm.git@3070dcc31948f48f9d8639b606af7bc019494fac'

from mlx_vlm import load, generate
from mlx_vlm.generate import generate_audio
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("nativ-community/LFM3-Audio-2B-4bit", trust_remote_code=False)
messages = [{"role": "system", "content": "Perform ASR in English."},
            {"role": "user", "content": ""}]
prompt = apply_chat_template(processor, model.config, messages, num_audios=1)
print(generate_audio(model, processor, prompt, audio=["speech.wav"], temperature=0,
                     generation_mode="text").text)

messages = [{"role": "system", "content": "Perform TTS in American English using a female voice."},
            {"role": "user", "content": "Hello there!"}]
prompt = apply_chat_template(processor, model.config, messages)
generate_audio(model, processor, prompt, output_audio_path="tts.wav", max_tokens=512)

Release validation

Strict native reload and short capability smoke checks on Apple M3 Ultra, MLX 0.32.4. Speech uses seed 42 and the model's default audio sampling settings. These fixtures are not a broader quality benchmark. Round-trip ASR uses the same model and is not an independent speech evaluation.

Check	Result
finite_logits	All allowed logits finite; speech placeholders correctly suppressed
arithmetic	-42 (expected 42; exact match: False)
asr	The quick brown fox jumps over the lazy dog. (WER 0.0%)
tts	Finite 24000 Hz waveform, 3.52 s; finish: stop
tts_roundtrip	The quick brown fox jumps over the lazy dog. (WER 0.0%)
interleaved	Finite 24000 Hz waveform, 2.32 s; finish: stop

Exact checks and settings are in validation.json; provenance and tensor audit are in conversion.json; file hashes are in SHA256SUMS. Generated speech examples are in samples/.

Implementation limits and terms

Speech generation supports batch one. Streaming waveform input/output and waveform-based speech history are not implemented. Video soundtrack input must be supplied separately. Omni thinking is unsupported. Lower precisions can reduce quality; each format needs task-specific evaluation before deployment.

Original checkpoint terms apply. See ORIGINAL_MODEL_CARD.md. No new weight license is assigned by this conversion.