Hibiki 1b Mlx Q4

--- license: cc-by-nc-sa-4.0 base_model: kyutai/hibiki-1b-mlx-bf16 tags: - mlx - speech-translation - text-to-speech - audio - quantized --- # Hibiki-M 1B MLX q4 This is a 4-bit MLX quantization of [`kyutai/hibiki-1b-mlx-bf16`](https://huggingface.co/kyutai/hibiki-1b-mlx-bf16) at revision `b3d6291f3dcf7954e1a502e4d66f32e3556f17ae`. The language-model weights were quantized with MLX `nn.quantize(bits=4, group_size=32)`, leaving tiny conditioner tensors dense when their last dimension is not divisible by 32. Mimi, tokenizer, and model config files are copied from the source repository. The staged `config.json` changes `moshi_name` to `hibiki-mlx-dc2cf5a5@80.q4.safetensors` so compatible loaders can apply q4 group-size-32 loading before strict weight load. ## Files - `hibiki-mlx-dc2cf5a5@80.q4.safetensors`: quantized Hibiki-M language-model weights - `mimi-dbaa9758@125.safetensors`: source Mimi codec weights - `tokenizer_spm_48k_multi6_2.model`: source sentencepiece tokenizer - `config.json`: source architecture config with q4 `moshi_name` This derivative keeps the upstream non-commercial share-alike license.

Source context: 7 downloads · 0 likes · Pipeline text-to-speech · Library mlx · Repo huybik/hibiki-1b-mlx-q4