coder543/ming-3.0-tts-tiny-glade

Ming 3.0 TTS Tiny for Glade

Hybrid conversion of Ming 3.0 TTS Tiny. The 7.9B-parameter mixture-of-experts language backbone runs through Swift MLX on the GPU; Core AI handles audio aggregation, flow inference and AudioVAE. The complete checkpoint contains 8,980,997,932 learned parameters before quantization, excluding registered buffers and the separately distributed CAMP+ speaker encoder. Output is 44.1 kHz mono. Despite the name, this is a large model: this conversion is qualified on a 16 GB M3 MacBook Air.

Using the bundle

Download one repository snapshot and keep the language, audio and speaker directories together. The optional GladeMingTiny product exposes MingTinyModelBundle and MingTinySynthesizer, including ordinary-text requests, the upstream task prompts and the common ordered TTS session interface. Python is not an application dependency. Core AI requires macOS 27 or iOS 27; iPhone execution and memory fit are not qualified here.

hf download coder543/ming-3.0-tts-tiny-glade --local-dir ./ming-tiny
glade tts --model ./ming-tiny --text "Welcome to the glade." \
  --voice US_James --output speech.wav

The built-in Chinese/English presets require no reference recording. Ordinary prompted speech instead requires reference audio and its matching transcript. Speaker-only controls, voice design, free-form instructions, paralinguistic speech and ordered multi-speaker scripts retain their separate upstream task contracts. Seven language controls are supported: Chinese, English, German, Spanish, French, Korean and Japanese. No voice recording is bundled here.

Each request produces a complete upstream utterance, preserving its silence. Online delivery orders completed requests; it does not stream unfinished latent frames. The default phased policy releases language weights before waveform rendering. Resident execution is also available. Memory mapping lets the OS manage file-backed pages; it does not remove the language model's working set.

Representation and runtime contract

• language/: aligned, memory-mapped language weights. Routed experts use affine 4-bit groups of 64. Other learned matrices retain BF16; recurrent state, router weights and decay parameters use FP32. The unused vocabulary head is omitted because speech consumes final hidden states. No additional training is applied.
• audio/: INT8-weight/FP16-activation Core AI flow and aggregation graphs, AudioVAE reference encoder/decoder graphs and host reconstruction resources.
• speaker/: the original CAMP+ weights and native MLX network description, with the source Kaldi filterbank and full-reference pooling conventions.
• The 24 language layers alternate three Kimi Delta Attention layers with one multi-latent attention layer. The first feed-forward layer is dense; later layers select eight of 128 routed experts and add a shared expert.
• Prompt and generated positions share a 4,096-position budget. The original old-shift generation loop retains its terminal latent and emits 13–250 latents at 12.5 Hz. The common interface reports an unfinished capped utterance as a runout rather than presenting it as complete speech.
• The sampler uses ten EPSS Euler steps, guidance unconditional + 2 * (conditional - unconditional), sigma 0.25, temperature 0 and sway -1. Noise/state are FP32; velocity, guidance and time increments reproduce the original BF16 rounding boundaries. Metadata records the complete sampler contract and default seeds.

Measured Mac performance

M3 MacBook Air, 16 GB, macOS 27.0.1; Release Glade runtime, US_James, seed 1234, phased execution. The complete ordinary-text input is:

Sunlight pours through the trees. A gentle breeze moves across the clearing. Welcome to Glade.

One same-text warmup precedes three measured passes. Initial preparation and WAV writing are excluded; recurring phase loads, text preparation, language/flow generation and waveform reconstruction are included.

Pass	Audio	Generation	RTFx
1	6.72 s	14.895 s	0.451×
2	6.72 s	18.717 s	0.359×
3	6.72 s	13.535 s	0.497×
Median	6.72 s	14.895 s	0.451×

A separate hardware trace confirms GPU language prefill/decode and ANE execution for all 840 flow, 83 aggregation and 22 waveform-decoder calls. No GPU intervals appear within those audio graph scopes. The capture is excluded from timing. ANE attribution uses exact call counts and temporal containment; these events do not measure arithmetic or bandwidth utilization.

The mapped language file is 5.8 GB (decimal). MLX reports peaks of 5.9 GB active and 0.9 GB cached allocations, sampled independently. These cannot be added into a measured simultaneous peak. The process physical-footprint plus Neural Engine no-footprint ledger peaks at 1.4 GB, but does not cover the mapped GPU model's complete memory requirement.

Qualification and limits

Numerical controls cover original tasks/tokenization, speaker features/network output, KDA/MLA recurrent states, flow sampling, reference conditioning and waveform reconstruction. Original-code controls use the exact packed expert values to distinguish quantization from native implementation differences. Retained final-condition RMS differences are 3.31–12.42% across the examined language steps/prefill partitions. Discrete expert routing can amplify small arithmetic differences; internal RMS is not a perceptual speech-quality score.

Cohere Transcribe recovers every source word from the short native English output. Repeated runs produce identical PCM, so they are one distinct audio check. Separate reference-conditioned checks recover the intended English, Chinese and German text. Reference choice matters: one German sentence with an English synthetic reference stops early in both Glade and the original BF16 implementation, while other tested references complete it. These checks do not qualify every voice/language combination or long-form generation. No smartphone performance or memory-fit claim is made.

Attribution

Checkpoint revision: b170071302cb0b06435ea784787b119f40a08be8. Source implementation: inclusionAI/Ming-3.0-tts, revision e999c39fa676722aaa9ea29b5aa19a90525b271f. Weights and source use MIT. LICENSE, NOTICE and CODE-LICENSE retain the Ming, FLA, CAMP+/3D-Speaker, TorchAudio and Kaldi credits and license texts. Core AI assets omit authoring debug locations while preserving graph signatures and operation statistics. Device specialization caches are not distributed.