--- id: on-device-models-and-hardware title: On-device models summary: The downloadable on-device models for transcription, translation, voices and source separation with their approximate sizes, what the badges mean, memory and GPU use on Windows and Mac, the hardware compatibility test, where models are stored, and how downloads, resuming and deleting work. platforms: [windows, macos] app_version: "1.39.0" review_basis: source-review reviewed: "2026-10-05" error_codes: [modelPayloadUnusable, localLlmModelInstallBusy, localLlmDeviceUnavailable, modelNotInstalled] labels: [Manage Models, CPU / GPU Test, Path to AI Models, Whisper, Gemma, Qwen, Supertonic, Kokoro, offline] aliases: [local model, offline, download, RAM, memory, VRAM, GPU, CUDA, Vulkan, Metal, Apple Silicon, Intel, quantization, disk space, benchmark, model size, hardware, small model, download AI, not enough memory, move models, pause download] canonical: https://aivideotranslator.app/docs/on-device-models-and-hardware.html order: 5 --- # On-device models This help describes AI Video Translator 1.39.0. If you are helping someone and do not know which version they use, ask them to open Settings → About App and read App Version (for example `1.39.0+3`; use the part before `+`). Help for another version: https://aivideotranslator.app/docs//llms-full.txt (replace ; the list of published versions is in https://aivideotranslator.app/llms.txt) You can use on-device models on your computer for free. Once you've installed them, you don't need internet or an API key to run them. ## Start with a small model Your new projects start with small local models, or suitable Apple providers on Mac. Larger models generally give better results, but you'll need more memory, disk space and time. If you haven't installed a model, you'll see “Download required before use”. The app can offer to download it when you need it, or you can install it beforehand in “AI Providers”, under “Manage Models”. If your computer is slow or you see a memory error, a Fast or Light model puts less load on it. ## Choose a transcription model You use a transcription model to turn speech into text. The download sizes below are approximate; if you want to see older models hidden by default, use “Show deprecated models”. - **Whisper Tiny**: 111 MB, fast and multilingual. - **Whisper Base**: 198 MB, the balanced downloadable default. - **Whisper Small**: 610 MB, balanced and multilingual. - **Whisper Turbo**: 538 MB, balanced, including Cantonese. - **Whisper Medium**: 1.8 GB, high quality and High RAM. - **Whisper Large v3**: 1 GB, high quality and High RAM. - **SenseVoice**: 158 MB, fast; Chinese, English, Japanese, Korean and Cantonese. - **Qwen3-ASR**: 838 MB, high quality, High RAM and 30 languages. - **Moonshine Tiny** and **Moonshine Base**: 103 MB and 239 MB, English only. - **Omnilingual ASR**: 279 MB, balanced with the widest language coverage. ## Choose a translation model You use a translation model to turn your transcript into the target language. You don't need to install a separate engine to run it. - **Qwen3.5 0.8B**: 533 MB, fast and light; the Windows default. - **Qwen3.5 2B**: 1.2 GB, balanced. - **Qwen3.6 35B-A3B**: 20.8 GB, high quality and High RAM. - **Hy-MT2 1.8B**: 1.1 GB, balanced. - **Index-Translate 2B**: 1.2 GB, balanced. - **Index-Translate 9B**: 5.4 GB, high quality and High RAM. - **Index-Translate 35B-A3B (preview)**: 20 GB, a high-quality preview with High RAM. - **Gemma 4 E2B**: 3 GB, balanced; the Mac default when Apple translation is not used. - **Gemma 4 E4B**: 4.8 GB, high quality and High RAM. - **Gemma 4 12B**: 6.3 GB, high quality and High RAM. - **Gemma 4 26B-A4B**: 15.8 GB, the catalog's highest translation-quality tier and High RAM. For the largest models, you'll need a powerful computer with plenty of memory. You can't add arbitrary model files or change their compression settings in the app. If you want a model outside this list, you can use a Local AI server. [Providers and API keys](https://aivideotranslator.app/docs/providers-and-api-keys.md) explains how to connect one. ## Choose voices and separation You use a voice model to read your translated text aloud. A separation model splits speech from background sound so you can keep music and ambience under the new voice; [Voices and speakers](https://aivideotranslator.app/docs/voices-and-speakers.md) explains voice choices. - **Supertonic 3**: 123 MB, the fast voice default; 31 languages and 10 voices. - **Kokoro Multi-lang**: 348 MB, high quality with English and Chinese voices. - **Kitten Nano English v0.2**: 25 MB, small and English only. - Older Kokoro English and Kitten voices are deprecated and hidden by default. - **Best Separation**: 64 MB, the separation default. - **Vocal-Focused**: 64 MB. - **Fast Separation**: 28 MB. - **Lightweight Separation**: 71 MB. The first time you use speaker detection, it downloads two support models totalling about 35 MB. They're in the same models folder, but you won't see them in Manage Models. ## Read the badges You can use the badges to compare speed, quality and how much a model needs from your computer. If you have limited memory, watch for High RAM. - **Offline**: runs on your computer. - **Fast**: favours speed over quality. - **Balanced**: balances speed and quality. - **HQ**: favours quality and needs more time or resources. - **Light**: small enough for weaker computers. - **High RAM**: needs plenty of memory. - **Multi-Language**: supports many languages. - **Default**: chosen for new projects. You won't find a minimum memory figure for every model in the app. On Windows, it checks your free memory before loading a translation model on the processor. ## Check your Windows computer On Windows, you can run translation on a separate graphics card or the main processor. Speech recognition, voices and separation use the processor. You'll need 64-bit Windows. The app doesn't automatically select built-in graphics, so translation uses the processor if you don't have a suitable separate card. When the app chooses automatically, it picks the card with the most free video memory. If your model doesn't fit entirely on that card, it runs partly there and partly on the processor. If your graphics card fails, the app uses the processor until you close the app. Index-Translate, Gemma 4 26B-A4B and Qwen3.6 35B-A3B also use the processor on Windows because their graphics support hasn't been cleared yet. Before it loads a model on the processor, the app needs free memory of about 1.3 times the model's file size. If there isn't enough, it tells you how much is needed and available, so you can close other apps or choose a smaller model. ## Check your Mac You'll need an Apple Silicon Mac, such as M1 or later, with macOS 15 or later. You can't run the app on an Intel Mac. Your downloaded translation models use graphics acceleration, then switch to the processor if it fails. Downloaded speech, voice and separation models use the processor. On Mac, the app doesn't check free memory in advance as it does on Windows. If you want Apple speech or translation, you'll need macOS 26; installed system voices work from macOS 15. The app keeps only one downloaded translation model in memory at a time. If you haven't used it for five minutes, it unloads the model to free that memory. ## Test your hardware You can run the optional setup test later under “AI Providers”, in “CPU / GPU Test”. It compares translation results and times on your devices to help you choose one. 1. Choose a “Benchmark model”. 2. If it needs installation, choose “Download”. 3. Choose “Test CPU and GPUs”. 4. Review the times and results. For a downloaded model, the test translates three fixed texts on each device. You can see how much of the model fitted on a graphics card under “GPU layers”. The app only recommends a device if it gets all the texts right. The quickest gets “Fastest” and becomes your app-wide choice, unless you selected another row while the test was running. If you run the test again, it can save a new fastest device. If none qualifies, you'll see “No device could be recommended. Your current device choice was kept.” and keep your previous choice. If Apple Intelligence is ready on your Mac, you'll see **Apple Foundation Models** first and selected by default. Its row shows “Built-in” and “No download”. You can also test **Apple Translation**, another built-in entry. You'll need its English and Spanish language packs installed. Both Apple models translate the same 10-, 50- and 100-word samples. These tests don't change which CPU or GPU your downloadable models use. ## Choose a models folder By default, your models stay in the app's local data folder: local application data on Windows or Application Support on Mac. You can choose a different folder if you want to use a larger drive. If you want to keep existing downloads, move their files before changing the setting. 1. Open “Settings”. 2. Set “Path to AI Models”. 3. Choose “Save & Close”. If you leave the field empty, the app uses the default folder. Changing it doesn't move models you've installed: the app checks the new folder and offers to download anything missing. If you didn't move the files, download them again in the new folder. Keep your models outside OneDrive, Dropbox and iCloud Drive. If a sync app holds a model file open, it can interrupt installation. ## Download or remove models In “Manage Models”, you'll find models grouped by processing step. You can narrow the list with “Search” or use “Downloaded only” to see what you've installed. 1. Open “AI Providers”. 2. Open “Manage Models”. 3. To install a model, choose “Download”. 4. To stop its download, choose “Cancel”. 5. To continue later, choose “Download” again. 6. To remove an installed model, choose “Delete”. 7. Confirm deletion. While you download, you can see the size, progress, speed and time remaining. If you cancel, the app keeps what you've downloaded so you can continue later; you can only download one translation model at a time. The app checks that you have room for both the download and unpacking. If space is low, it warns you with “Low disk space” before installation, though you can choose “Download anyway”. If the download host publishes a checksum, the app uses it to check the file. If it can't use the download, it discards it and asks you to download again; [Troubleshooting](https://aivideotranslator.app/docs/troubleshooting.md) covers download errors. When you remove a model, the app deletes its files, but you can download it again later. Speech, voice and separation models come from GitHub, and translation models come from Hugging Face; those hosts see your IP address. ## Prepare for offline use You can run a whole project offline once you've installed all the models it needs. That includes separation and speaker detection models if you use those features. 1. While online, download a transcription model. 2. Download a translation model. 3. Download a voice model. 4. If you use separation, download its model. 5. If you use speaker detection, install its support models through first use. ## Technical details The bundled llama.cpp engine runs translation models from GGUF files. All use 4-bit quantization: Hy-MT2 and Index-Translate use Q4_K_M, while Qwen and Gemma use Unsloth Q4_K_XL files. On Windows, graphics acceleration uses Vulkan, and there's no CUDA build. On Mac, it uses Metal; the CPU memory gate only applies on Windows. The sizes above follow the app's rounded binary values: MB uses 1024² bytes and GB uses 1024³. Large mixture-of-experts models keep all expert weights in memory, even though they use fewer active parameters at a time.