groxaxo/Chatterbox-LatAm-MLX

Chatterbox LatAm MLX weights

Apple Silicon MLX conversions of ResembleAI/Chatterbox-Multilingual-es-mx-latam, with a full FP32 baseline, FP16 T3 candidate, and oMLX Q8/Q6/Q5 T3 variants. Each variant is in its own models/ directory. The required speech tokenizer is included at speech-tokenizer/.

Directory	T3 precision	Remaining model	Model weights
models/oQ8	210 attention/MLP matrices at Q8	FP32	1.264 GB
models/FP16T3	All 292 T3 tensors at FP16	FP32	1.639 GB
models/FP32	FP32	FP32	2.711 GB
models/oQ6	210 attention/MLP matrices at Q6	FP32	1.138 GB
models/oQ5	196 matrices at Q5 and 14 at Q6	FP32	1.080 GB

The speech tokenizer adds 0.495 GB. Sizes use decimal GB. The complete set of weights is about 8.33 GB. Q variants use the structural oMLX policy without speech-activation calibration. Per-layer quantization settings are preserved in each variant's config.json.

Download

The repository is public and does not require login. For Q8 and its tokenizer:

from huggingface_hub import snapshot_download
snapshot_download(
    "groxaxo/Chatterbox-LatAm-MLX",
    allow_patterns=["models/oQ8/*", "speech-tokenizer/*"],
    local_dir="chatterbox-latam-mlx",
    token=False,
)

Remove allow_patterns to download all variants and notices.

Runtime compatibility

The fixed runner and Talk adapter are public. Follow that repository's installer and Talk integration guide to use the corrected vocoder and saturation guard.

Use a Chatterbox-compatible MLX runner that supports the per-layer quantization settings. For runtime parity with the original V3 vocoder, the final vocoder LeakyReLU negative slope must be 0.01; intermediate slopes remain 0.1. Using 0.1 for the final activation can produce saturated audio. The weights do not contain this runtime correction.

The FP16 candidate changes stored T3 weights only. Activation and KV-cache precision depend on the runner. Quantization and precision changes should be evaluated for the intended text, reference voice, and runtime.

Source and notices

LatAm source revision: 27e595bf2fe7be0533ca299d9afafcde08b7cca7. Speech tokenizer: mlx-community/S3TokenizerV2, revision e0c9886f0e1c35ae85b1f27277416fb19fc72bec.

The original Resemble AI MIT license is retained in LICENSE. Dependency provenance is in NOTICE.md. SHA256SUMS.json records artifact hashes.