mlx-q6-higgs-tts-3-4b

mlx-q6-higgs-tts-3-4b

Derived from Higgs TTS 3, licensed from Boson AI USA, Inc.

This model was converted to MLX format from bosonai/higgs-tts-3-4b using mlx-audio 0.4.5 (6-bit, group_size 64, 6.550 bpw). Refer to the original model card for more details.

License

Research and non-commercial use only, per the included LICENSE (Boson Higgs TTS 3 Research and Non-Commercial License). Commercial use requires a separate license from Boson AI ([email protected]). The Boson Acceptable Use Policy applies. See the NOTICE file for attribution and the list of modifications.

Requirements

• mlx-audio ≥ 0.4.5
• Apple Silicon with 8 GB unified memory is enough (~3.7 GB of weights) — works even on the entry-level MacBook Neo (A18 Pro, 8 GB)
• model_type in config.json is already patched to higgs_audio_v3 — without this patch, model detection in mlx-audio 0.4.5 is non-deterministic (see Blaizzy/mlx-audio#854)

What was changed

• Quantized to 6-bit (group_size 64, affine, 6.550 bpw) with the official mlx_audio.convert from mlx-audio 0.4.5.
• Audio codec weights extracted from the LM checkpoint into codec.safetensors.
• model.safetensors.index.json removed (required for mlx-audio codec discovery).
• config.json model_type set to higgs_audio_v3 (was higgs_multimodal_qwen3).

See NOTICE for the full modification list.

Performance

Probe benchmark on a MacBook Neo (A18 Pro, 8 GB unified memory), mlx-audio 0.4.5:

Model load: ~4.5 s. The benchmark ran with memory pressure from other apps (~3 GB used, swap active), so these RTF numbers may be slightly pessimistic.

Listening evaluation: quality is good — the first usable MLX quant of Higgs v3. A 4-bit quant (4.56 bpw) of the same model was rejected in listening tests.

Samples

Generated with the default voice from this repository's weights (downloaded back from the Hub, so they also serve as an end-to-end integrity check). The English sample is the exact command from Usage below.

English — “Hello from the 6-bit Higgs TTS.”

<audio controls src="https://huggingface.co/whitelabel/mlx-q6-higgs-tts-3-4b/resolve/main/samples/hello_en.wav"></audio>

Russian — «Привет! Это шестибитный квант Higgs TTS, запущенный на MacBook Neo.»

<audio controls src="https://huggingface.co/whitelabel/mlx-q6-higgs-tts-3-4b/resolve/main/samples/privet_ru.wav"></audio>

Usage

python -m mlx_audio.tts.generate --model whitelabel/mlx-q6-higgs-tts-3-4b \
  --text "Hello from the 6-bit Higgs TTS." --file_prefix hello

The output is written as hello_000.wav to the current directory (use --output_path DIR to change it). For voice cloning, pass --ref_audio ref.wav and --ref_text "transcript of the reference".