1. Third-Party Models Are Not Bundled
The app does not bundle third-party model weight files by default. Recommended models are presented as download entries. A model file is accessed only when the user actively chooses to download or import it.
2. Current Recommended Sources
The current recommended model entries point to public ModelScope pages or files, including Qwen, Gemma, NVIDIA Nemotron, DeepSeek, and Phi families. Model names, file names, licenses, availability, download speed, and later changes are controlled by the corresponding model publisher or platform.
- Qwen3.5 4B: ModelScope · Q4_0 GGUF · the public model page identifies Apache-2.0 as the displayed license · model file page
- Gemma 4 E4B: ModelScope · Q4_0 GGUF · the public model page identifies Apache-2.0 as the displayed license · model file page
- Gemma 4 E2B: ModelScope · Q4_0 GGUF · the public model page identifies Apache-2.0 as the displayed license · model file page
- Qwen3.5 9B: ModelScope · Q4_0 GGUF · the public model page identifies Apache-2.0 as the displayed license · model file page
- Qwen3.5 2B: ModelScope · Q4_K_S GGUF · the public model page identifies Apache-2.0 as the displayed license · model file page
- NVIDIA Nemotron 3 Nano 4B: ModelScope · Q4_0 GGUF · the public model page identifies NVIDIA Nemotron Open Model License / Other; users should review the full terms separately · model file page
- DeepSeek R1 Distill Qwen 7B: ModelScope · Q4_K_M GGUF · public pages may refer to MIT / Apache-2.0 related terms; users should rely on the model page and full license text · model file page
- Phi-4 mini instruct: ModelScope · Q4_K_M GGUF · the public model page identifies MIT as the displayed license · model file page
The license notes above are provided only as release-preparation references based on public model pages available for this version. They are not a legal opinion or a license guarantee. Model publishers, hosting platforms, and file contents may change later, so users should verify the model page and full license text before downloading, running, or using a model commercially.
The app does not preinstall or redistribute the model weights listed above inside the app bundle. Recommended downloads are initiated by the user and saved to a local model folder selected through the system file picker. If a third-party platform, model publisher, or model license requires sign-in, additional acceptance, attribution, use restrictions, regional restrictions, or other conditions, users are responsible for confirming and following those conditions.
3. Current Recommended Whisper Speech Model Entries
The audio transcription feature may provide Whisper ggml / ggml-q5_0 speech model download entries. These model files are not preinstalled in the app bundle. The app accesses a third-party link only when the user actively starts a download, and saves the file into the app sandbox location used for speech models.
- Whisper Tiny: ModelScope ·
ggml-tiny.bin· about 75MiB · model file page - Whisper Base: ModelScope ·
ggml-base.bin· about 142MiB · model file page - Whisper Small: ModelScope ·
ggml-small.bin· about 466MiB · model file page - Whisper Medium: ModelScope ·
ggml-medium.bin· about 1.5GiB · model file page - Whisper Large v3 Turbo: ModelScope ·
ggml-large-v3-turbo-q5_0.bin· about 547MiB · model file page
Whisper model files are static model weight / tensor data files used for local speech recognition. The app reads those files through the bundled local whisper.cpp / ggml runtime submitted with the app for review. Recommended entries are only convenient access points for user-initiated downloads of publicly available model files, and do not indicate official cooperation, authorization, endorsement, agency, distribution, affiliation, or joint operation with OpenAI, Whisper, ModelScope, the mirror maintainer, or any related entity.
4. Nature of GGUF and Whisper Model Files
Recommended chat model entries are usually .gguf model weight / tensor data files that are read by the app's bundled local llama.cpp / ggml runtime for inference. Recommended Whisper entries are usually .bin model weight / tensor data files that are read by the app's bundled local whisper.cpp / ggml runtime for speech recognition. These files are not iOS executables, scripts, dynamic libraries, plug-ins, extensions, code packages, or installable apps. The app does not download and execute new binary code from third-party platforms; the native inference and speech-recognition runtimes are submitted with the app for review.
After download, the app performs basic local validation such as extension, file size, GGUF header, or Whisper model filename checks. These checks help identify compatibility problems and obvious file errors, but they are not a security audit, license review, source certification, or complete hash verification.
5. No Affiliation or Endorsement
Listing a model source does not mean the app, developer, model publisher, model platform, or trademark owner has an official partnership, authorization, endorsement, agency, distribution, affiliation, or joint operation relationship.
6. Brand Names and Neutral Badges
The app may reference names such as Qwen, Gemma, DeepSeek, Phi, NVIDIA, GPT-OSS, Whisper, OpenAI, and ModelScope only to identify model families, model sources, or compatible formats. The App Store release does not use official logos from those third-party models or platforms. Visual model marks in the app are app-owned neutral text badges and do not indicate official cooperation, authorization, endorsement, approval, or joint operation. The relevant names, trademarks, model names, and platform names belong to their respective owners.
7. User Responsibilities
Before downloading, importing, or running a third-party chat model or speech model, users should read and comply with the model page, platform terms, model license, acceptable use policy, and applicable laws. Users are responsible for their chosen model and use case.
8. Output and Transcription Risks
Third-party chat model output may contain errors, outdated information, hallucinations, bias, misleading statements, offensive content, incomplete information, or content unsuitable for a particular purpose. Whisper speech model transcripts may contain missing words, wrong words, sentence-boundary errors, punctuation errors, language-detection errors, timestamp drift, or complete recognition failure. Users should judge the accuracy, legality, safety, and suitability of outputs and transcripts themselves.
9. Platform Data Handling
When users visit a third-party platform or download a model, the platform may process network request information under its own privacy policy and terms. This app does not control third-party servers, domains, rules, licenses, file availability, or data handling.