Third-party notices
PUSHTYPE is built on open-source software and open speech models. We're grateful to the people who made them. This file lists each one, its licence and what the licence asks of us. The full licence texts are in legal/third-party/, and the release build copies this whole folder into PUSH.app/Contents/Resources/Legal.
The companies named here have not endorsed PUSHTYPE. Their names are used only to say where each component comes from.
Software shipped inside PUSH.app
These are compiled into the app or bundled with it. Their licences require us to include the copyright notice and licence text with every copy we distribute.
| Component | Version | Licence | Copyright | Full text |
|---|---|---|---|---|
| FluidAudio | 0.17.4 | Apache License 2.0 | FluidInference | third-party/FluidAudio/LICENSE.txt |
| ↳ NemoTextProcessing (text-processing-rs), linked by FluidAudio | 0.3.1 | Apache License 2.0 (with MIT/Apache sub-components) | FluidInference and contributors | third-party/FluidAudio/NemoTextProcessing-LICENSE.md |
| ↳ fastcluster, compiled into FluidAudio | — | BSD 2-Clause | © 2011 Daniel Müllner; © Google Inc. | third-party/FluidAudio/fastcluster-LICENSE.md |
| ↳ VBx, ported in FluidAudio | — | Apache License 2.0 | VBx authors | third-party/FluidAudio/vbx-LICENSE.md |
| ↳ Kokoro text frontends, in FluidAudio | — | Apache 2.0 / MIT / BSD (see file) | Various | third-party/FluidAudio/JapaneseG2P-LICENSE.md, KokoroAneSpanishFrenchG2P-LICENSE.md |
| Sparkle | 2.9.3 | MIT, plus BSD-style and public-domain components (bsdiff, sais-lite, ed25519, SUSignatureVerifier) | Andy Matuschak and the Sparkle contributors | third-party/Sparkle-LICENSE.txt |
| LaunchAtLogin-Modern | 1.1.0 | MIT | Sindre Sorhus | third-party/LaunchAtLogin-Modern-LICENSE.txt |
voice-glow, ported as VoiceGlow.swift | — | MIT | © 2026 Jakub Antalik | third-party/voice-glow-LICENSE.txt |
| llama.cpp (with ggml), Type's text engine | b11147 | MIT | © 2023-2026 The ggml authors | third-party/llama.cpp-LICENSE.txt |
We didn't change FluidAudio, Sparkle, LaunchAtLogin or llama.cpp. They're used as released (llama.cpp as its official prebuilt framework). Apache 2.0 §4(b) asks that modified files be marked; since we modified nothing, there is nothing to mark.
Speech models (downloaded on first use, not bundled)
PUSHTYPE doesn't ship model weights. When you press Download in Settings, PUSHTYPE fetches the model you picked straight from its public Hugging Face repository onto your Mac. PUSHTYPE never sends your audio anywhere; the models run on your Mac.
We don't redistribute the weights, but we credit them anyway: it's what their authors ask for, and CC BY 4.0 requires it when a work is shared.
| In PUSHTYPE | Model | Made by | Licence | Notice |
|---|---|---|---|---|
| Parakeet Ultra (default) | moondream/parakeet-ultra, a post-training of nvidia/parakeet-tdt-0.6b-v3; Core ML conversion FluidInference/parakeet-ultra-coreml | Moondream, from NVIDIA's Parakeet TDT 0.6B v3; converted by FluidInference | CC BY 4.0 | "Parakeet Ultra" by Moondream, based on "Parakeet TDT 0.6B v3" by NVIDIA, licensed under CC BY 4.0; converted to Core ML by FluidInference. |
| Parakeet Unified, Parakeet Streaming | nvidia/parakeet-unified-en-0.6b; Core ML conversion FluidInference/parakeet-unified-en-0.6b-coreml | NVIDIA; converted by FluidInference | NVIDIA Open Model License | Licensed by NVIDIA Corporation under the NVIDIA Open Model License. |
| Nemotron Multilingual | NVIDIA Nemotron 3.5 ASR Streaming Multilingual 0.6B; Core ML conversion FluidInference/Nemotron-3.5-ASR-Streaming-Multilingual-0.6b-CoreML | NVIDIA; converted by FluidInference | OpenMDW-1.1 | NVIDIA Nemotron 3.5 ASR Streaming Multilingual, © NVIDIA Corporation, made available under OpenMDW-1.1. |
| Voice activity detection | Silero VAD; Core ML conversion FluidInference/silero-vad-coreml | Silero Team; converted by FluidInference | MIT | Copyright (c) 2020-present Silero Team. |
Before release: confirm each model licence on its Hugging Face card. We compiled the table above from FluidAudio's source and documentation plus public listings of the upstream models. Hugging Face was unreachable from the environment this file was written in. Before the first paid release, open each model card, check the licence field and copy any
NOTICEtext word for word. If you find a mismatch, fix it here and in the website's legal page.
Typing models (downloaded on first use, not bundled)
Type, the inline autocomplete, runs one of these text models on your Mac through llama.cpp. As with the speech models, PUSHTYPE downloads the file you pick from Hugging Face; we don't redistribute the weights. Each is the quantized GGUF build by mradermacher of the original model, and every one is licensed under the Apache License 2.0 (checked on each model card, 2026-10-12).
| In PUSHTYPE | Model | Made by | GGUF build | Licence |
|---|---|---|---|---|
| Recommended | Qwen/Qwen3.5-2B-Base | Qwen team, Alibaba Cloud | mradermacher/Qwen3.5-2B-Base-i1-GGUF | Apache 2.0 |
| Best | google/gemma-4-E2B | mradermacher/gemma-4-E2B-i1-GGUF | Apache 2.0 | |
| Pro | google/gemma-4-E4B | mradermacher/gemma-4-E4B-i1-GGUF | Apache 2.0 |
All four speech-model licences allow commercial use. What each one requires:
- CC BY 4.0. Credit the creators, link to the licence, say whether changes were made (we made none) and don't imply endorsement.
- NVIDIA Open Model License. When redistributing, pass on the agreement and include the notice "Licensed by NVIDIA Corporation under the NVIDIA Open Model License". Rights end if you sue over the model's IP, or if you bypass its guardrails without a substitute. Using NVIDIA's trademarks isn't granted: we name NVIDIA only to say where the model comes from, and never use its logo.
- OpenMDW-1.1. When redistributing, keep the licence and the notices of origin. Rights end on patent or copyright litigation over the model. There are no restrictions on outputs, so your transcripts are yours.
- MIT (Silero VAD). Keep the copyright notice.