Rill Local Model Notices¶
Rill does not bundle model weights in the application. Models are downloaded only after the user enables them. Every production model is fetched at the exact revision listed below and every reviewed file is verified against the byte count and SHA-256 recorded in the Swift catalogs before it is published to Rill's local model directory.
Speech recognition¶
| Rill model ID | Upstream repository | Pinned revision |
|---|---|---|
qwen3-asr-0.6b-mlx-8bit |
mlx-community/Qwen3-ASR-0.6B-8bit |
89e96d92ba34aca20b3e29fb10cc284097d1219f |
qwen3-asr-1.7b-mlx-8bit |
mlx-community/Qwen3-ASR-1.7B-8bit |
a8379a2e2f9e313c9292cdf1af4055ab56d50d55 |
The exact reviewed file inventory is defined in Sources/RillSpeechContracts/LocalSpeechModelCatalog.swift. Qwen model repositories publish their license and model-card terms alongside the pinned weights. Rill uses these models locally through mlx-audio-swift; recognition audio is not uploaded by the local provider.
Voice activity detection¶
| Purpose | Upstream repository | Pinned revision |
|---|---|---|
| Streaming VAD | mlx-community/silero-vad-v6 |
2ebf4a5e10726a2e78ddd4d70eedfb6f1c33eb06 |
The reviewed model.safetensors and config.json hashes are defined in Sources/RillMLXRuntime/MLXSileroVADRuntime.swift. Silero VAD is distributed under the MIT license by its upstream authors. Rill downloads this fixed MLX revision on demand; no ONNX VAD is packaged in the application.
Speech synthesis¶
| Rill model ID | Upstream repository | Pinned revision |
|---|---|---|
qwen3-tts-0.6b-customvoice-mlx-4bit |
mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-4bit |
08c72cad5e2fd0f41730c8bd1f28149585e46361 |
qwen3-tts-0.6b-customvoice-mlx-8bit |
mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-8bit |
049ef77fe8816b536193c0c25f9a214d17921282 |
qwen3-tts-0.6b-customvoice-mlx-bf16 |
mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-bf16 |
6415d95f88be018ff9e46813119dc3bc12261328 |
The exact TTS file inventory is defined in Sources/RillSpeechContracts/SpeechSynthesisModelCatalog.swift. TTS is optional: its worker is started only for a resident TTS model or an active synthesis request.
Clipboard semantic search¶
| Rill model ID | Upstream repository | Pinned revision |
|---|---|---|
qwen3-embedding-0.6b |
Qwen/Qwen3-Embedding-0.6B |
97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3 |
The upstream model card declares Apache-2.0. RecordEmbeddingModelCatalog.swift pins the five required files, byte counts, and SHA-256 values (approximately 1.2 GB). Search downloads require the explicit panel action. MLXEmbedders runs the model locally; search queries and clipboard records are never model-download inputs. Weights are not distributed in the App bundle.
Runtime boundary¶
mlx-audio-swift and its SwiftPM dependency graph are pinned in Package.resolved and documented in THIRD_PARTY_NOTICES.md. The main application does not perform MLX inference; separate supervised ASR, TTS, and clipboard-search helper processes own their respective model caches.