> ## Documentation Index > Fetch the complete documentation index at: https://docs.vowen.ai/llms.txt > Use this file to discover all available pages before exploring further. # Model Download & Management > Download, manage and remove local speech models, and what Vowen downloads on your behalf. ## Your first model Vowen offers you a model to download during onboarding. | Platform | Offered | Default | | - | - | - | | macOS | **Parakeet V2** or **Parakeet V3** | Parakeet V2, marked Recommended | | Windows | **Parakeet V2** or **Parakeet V3** | Parakeet V2, marked Recommended | | Linux | Whisper models | **Base** | **Parakeet V2** is English-only and the fastest of the family. If you speak anything other than English, expand the **Parakeet V3** card in the same step and pick that instead: it auto-detects across 25 European languages. You can change your mind at any time from the **Speech models** page in the sidebar. ## Downloading more models Click **Speech models** in the left sidebar. Local models are listed under **English Only** and **Multilingual**. Each card shows the model's name and download size. A card you do not have yet carries a download arrow on the right. Click the card. The whole card is the button, and the download starts straight away, so there is no confirmation step to catch a mis-click. Progress replaces the logo with a ring. Clicking a model card you have **not** installed begins the download immediately. On Large v3 that is 1.5 GB. Check the size on the card before you click. A local model card mid-download, showing a circular progress ring with a cancel square The model is selected automatically once the download finishes. The Local section of the Speech models page listing model cards with their sizes and download buttons On corporate or restricted networks, in-app downloads can fail with a certificate error. See [Manual Model Installation](/troubleshooting/manual-model-install) for how to download models in your browser and install them yourself. ## First-run optimization The first time a Parakeet or Nemotron model runs on your Mac, it is optimized for your specific chip. This is a one-time cost per model, and it is why an otherwise instant model appears to hang the first time. * It takes roughly **20 seconds**, and noticeably longer the very first time you ever run one. * While it runs, the indicator shows **"Optimizing Parakeet model — first-run optimization for your device."** **A dictation you start during optimization is saved, not transcribed.** Vowen keeps the audio and adds a row to your [Voice Log](/transcription/voice-log) reading *"Transcription pending — model is optimizing"*. Once optimization finishes, open that entry's menu and choose **Retry transcript**. Nothing is lost, but the text will not appear in your app until you do. After the first run the model stays warm in a background service, so ordinary dictations pay no load cost at all. ## The hidden streaming model If you use Parakeet on macOS, there is a model in your models folder you never asked for. **Parakeet V2 and V3 need a separate end-of-utterance streaming model** to power the live transcription preview. It is about 200 MB, and Vowen downloads it quietly in the background after the batch model finishes. It is deliberately kept out of the visible download progress so the bar reflects the model you actually picked. * It is shared. V2 and V3 use the same weights, so having both installed does not download it twice. * It is **removed automatically** when you uninstall your last Parakeet model. * **Parakeet Japanese and Mandarin do not use it**, so installing only those never pulls it down. * It is macOS only. If you are auditing disk use and find a `parakeet-eou-streaming` folder you do not recognize, that is what it is. ## Other automatic downloads | What | When | Size | | - | - | - | | Parakeet end-of-utterance streaming model | After installing Parakeet V2 or V3 on macOS | \~200 MB | | Speaker diarization bundle (Windows) | The first time you use Identify Speakers on Windows | \~30 MB | | Speaker diarization models (macOS) | The first time you use Identify Speakers on macOS | Fetched by the speech engine | | Streaming voice activity detection (macOS) | The first time Parakeet streaming runs | A few MB, cached by the speech engine | None of these appear as selectable models, because none of them are. ## Managing models ### Switching models Click any downloaded model to make it active. You can also switch models from the tray icon's right-click menu without opening the app. ### Deleting models Click the delete icon on a downloaded model's card to free up disk space. You can re-download it later. The icon only appears on models you are **not** currently using, so the active model cannot be deleted by accident. To remove the one you are on, switch to another model first and the delete icon appears on the old one. ### Disk space | Model | Download | | - | - | | Whisper Tiny / Tiny (EN) | \~74 MB | | Whisper Base / Base (EN) | \~141 MB | | Whisper Small / Small (EN) | \~465 MB | | Whisper Medium / Medium (EN) | \~1.4 GB | | Whisper Large v3 | \~2.9 GB | | Whisper Large v3 Turbo | \~1.5 GB | | Parakeet V2 | \~430 MB (macOS) / \~451 MB (Windows) | | Parakeet V3 | \~445 MB (macOS) / \~478 MB (Windows) | | Parakeet Japanese (macOS) | \~532 MB | | Parakeet Mandarin (macOS) | \~1.5 GB | | Nemotron EN (macOS) | \~528 MB | | Nemotron Multilingual (macOS) | \~588 MB | Add roughly 200 MB on top if you have any Parakeet V2 or V3 installed on macOS, for the streaming model above. Parakeet ships a different build on each platform, which is why the download sizes differ. Parakeet Japanese, Parakeet Mandarin and both Nemotron models are macOS only and do not appear on the Windows list at all. ## GPU acceleration (Windows) Local Whisper transcription on Windows can be accelerated with an NVIDIA GPU. Requirements: * **Windows only.** Macs need nothing extra; Linux is not supported. * An **RTX 2000 series or newer** NVIDIA GPU. AMD and Intel GPUs are not supported. * About **631 MB** of disk for the CUDA module. You do **not** need to install the CUDA toolkit separately. Scroll to the bottom of the **Speech models** page. The **GPU Acceleration** section only appears when Vowen detects a supported GPU, so if you cannot see it, your card is not eligible. Click the download button on the **NVIDIA CUDA** card. The card shows progress in MB and can be canceled. The card then reads **GPU active · RTX 2000+**. You can remove the module later with the delete button next to it. With GPU acceleration, even Whisper Large v3 responds in a couple of seconds. ## Resource Efficient Mode On a constrained machine, enable **Resource Efficient Mode** at the bottom of **Settings > General**. Vowen then runs the model as a one-shot process instead of keeping it resident: lower memory use, slightly slower transcription. The row is hidden when it cannot do anything. That means with a cloud model selected, and with Parakeet on macOS, which has no one-shot path. ## Cloud models Cloud models need no download and use no local disk. Sign up with the provider, generate an API key, and paste it into Vowen. Open **Speech models** and choose one from the Cloud section. Use the link in the panel to open the provider's dashboard. Groq, Deepgram, AssemblyAI, Gemini and others have free tiers that cover ordinary use. Paste the key and save. The model is ready immediately. Connect as many providers as you like and switch between them at any time. See [Models & Engines](/transcription/models#cloud-models) for what each one offers. *** Have a question about downloading or connecting models? See the [Models FAQ](/faq#models). This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.