maya-go-macos

Weights for running Maya 2 Native (Indic + English TTS, 20 built-in voices, 24 kHz) in the Adsorp Go server on macOS Apple Silicon only (MLX / Metal). These are the unmodified upstream files from maya-research/maya2-native plus the kyutai/mimi codec, laid out for the Go server and listed in manifest.json (size + SHA-256 per file).

• model-0000x-of-00002.safetensors, model.safetensors.index.json, maya3_config.json, config.json — Maya 2 Native (bfloat16; the server quantises to 8-bit at load)
• q6/model.safetensors — the same model pre-quantised to 6-bit (group 64, ≈2.7 GB, produced by the Go server's --export-packed); used instead of the two shards above when 6-bit is chosen (faster, smaller download, close to 8-bit quality). manifest.json tags files group: full (the original shards, used for 8-bit) or group: q6; untagged files are shared. An older experimental q4/model.safetensors may still exist in this repo; it is no longer used because its voice quality was clearly worse.
• backbone/ — Qwen3 tokenizer and config
• voices/voices.json — the 20 named voices
• mimi/ — Mimi codec (CC-BY-4.0, Kyutai)
• duration_prior.json, text_frames_prior.json — length priors used by the stop guards

Needs roughly 7 GB of unified memory at 8-bit (6-bit: less, and faster). Not intended for Windows, Linux or Intel Macs. Voice cloning (ECAPA speaker encoder) is not included.

See NOTICE.md and LICENSE for attribution.