reiscook/pocket-tts-mlx

Pocket TTS - MLX Weights

MLX-converted weights for Kyutai's Pocket TTS model. Optimized for Apple Silicon inference via MLX (Python) and MLX-Swift.

Weight Variants

Directory	Size	Description	RTF (M2 Air)
bf16/	224MB	bfloat16 baseline	~3x realtime
int8/	148MB	8-bit quantized FlowLM, bf16 Mimi	~5x realtime
int4/	107MB	4-bit quantized FlowLM, bf16 Mimi	~6x realtime

All variants include FlowLM + Mimi decoder in a single unified mlx_model.safetensors file.

Voice Embeddings

8 pre-extracted voice embeddings from the Kyutai release:

• voice/alba.safetensors
• voice/azelma.safetensors
• voice/cosette.safetensors
• voice/eponine.safetensors
• voice/fantine.safetensors
• voice/javert.safetensors
• voice/jean.safetensors
• voice/marius.safetensors

See GitHub repo for full source code. (coming soon)

Source

Converted from kyutai/pocket-tts official weights.