# Which model should you use?. Dimmy Help. > Which model should you use?. Step-by-step help for Dimmy, the native voice dictation app for macOS, Windows, and Linux. Source: https://dimmy.app/help/choose-your-model/ [← Back to Help center](https://dimmy.app/help/) Providers & Models # Which model should you use? Four questions decide it: what machine you have, what must not leave it, which subscriptions you already pay for, and which language you speak. Answer them and the rest follows. tip If you just want it to work: install, keep the defaults, dictate. Dimmy already picks a sensible model for your machine. Come back here when you want to know why, or when something feels slow, wrong, or too chatty with the internet. ## Three choices, not one People say "which model" as if there were one. Dimmy actually asks you three separate questions, and you can answer them differently. | Stage | What it does | Can be | | -------------- | ------------------------------------------------- | ---------------------------------------------------------------------- | | Speech to text | Turns your voice into words. | Local engine, or a cloud provider. | | Rewrite | Applies a style or tone to those words. Optional. | Off, local LLM, cloud LLM, or your own Claude / ChatGPT / Gemini plan. | | Recap | Summarises a whole meeting. Optional. | Same options as rewrite, picked independently. | Nothing forces them to match. Local speech to text with a cloud recap is a perfectly normal setup, and it is the one most privacy-minded users land on. ## Question 1: what machine do you have Dimmy reads your hardware at first run and preselects for you. It never hides a model and never blocks one. The worst it does is preselect the cloud card, show you an amber or red dot next to a model that will not fit, and free up memory before loading something large. | Your machine | Dimmy's verdict | What it means | | --------------------------- | --------------- | -------------------------------------------------------------------------------------- | | Apple Silicon Mac | Good | Always. Unified memory plus the Neural Engine, which nothing else is fighting you for. | | Dedicated GPU, 4 GB or more | Good | Any local model on the list is realistic. | | Dedicated GPU, 2 to 4 GB | Tight | Smaller models only. Large ones will load, then crawl or fail. | | Integrated graphics | Poor | Cloud is a better start, no matter how much system RAM you have. | | GPU could not be read | Unknown | Treated as fine, not as weak. Dimmy will not push you to the cloud on a guess. | info The verdict answers whether a model fits, never how fast it will be. The same 4 GB card measured 8 seconds per window one morning and 2 seconds the same afternoon, because its power limit was stuck. Treat the dots as a floor, not a forecast. ## Question 2: what must not leave your computer This is the question worth answering honestly, because it is the one that is hard to undo. Here is exactly what goes out in each setup. | Setup | Audio leaves? | Text leaves? | Who receives it | | ------------------------------------ | ------------------------------ | ---------------------- | ---------------------------------------------------------------------------------------------------------- | | Local speech to text, no rewrite | No | No | Nobody. | | Local speech to text, local LLM | No | No | Nobody. | | Local speech to text, cloud rewrite | No | Yes, the transcription | Only the vendor whose model you picked. | | Cloud speech to text, cloud rewrite | Yes, the recording is uploaded | Yes | The speech vendor gets your audio, the rewrite vendor gets your text. They can be two different companies. | | A CLI bridge (Claude, Codex, Gemini) | No | Yes, the text | Your own plan, through the vendor's own official command. Dimmy never reads your credentials. | For a recap, the entire meeting transcript goes to whichever provider you picked for it. A local recap keeps all of it on the machine. Either way the audio, the transcript and the recap stay saved in your recordings folder. info Three things reach the network whatever you choose: the licence check, the update check, and telemetry unless you turn it off. Everything else, local speech to text, local rewrite, local recap, history and key storage, works with no connection at all. See [What Dimmy sends](https://dimmy.app/help/telemetry/). ## Question 3: which subscriptions do you already pay for This is the part people miss, and it is the one that saves the most money. If you already pay for Claude, ChatGPT or a Gemini Code Assist work seat, Dimmy can run the rewrite and the recap through that plan instead of an API key. It drives the vendor's own command line tool, so the work is billed against the plan you already have and costs no extra credit. | You already pay for | Turn on | Notes | | ------------------------------------------ | ------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Claude (Pro, Max, Team) | Claude CLI | See [Claude CLI setup](https://dimmy.app/help/integrations-claude-cli/). | | ChatGPT (Plus, Pro, Team) | Codex CLI | See [Codex CLI setup](https://dimmy.app/help/integrations-codex-cli/). | | Gemini Code Assist, Standard or Enterprise | Gemini CLI | Work accounts only. Google ended Gemini CLI access for personal accounts, including AI Pro and Ultra, so you have to confirm yours is a company seat before the option appears. | | Nothing yet | An API key, or stay local | See [Add your API key](https://dimmy.app/help/api-keys/). | warning Two different things are both called a subscription. Your Dimmy plan pays for signed, notarised builds and the in-app updater. It does not change which models you can use: every cloud model works with your own API key or your own provider plan, on any plan. The "Use subscription" toggle in Settings is the other sense, your Claude or ChatGPT or Gemini plan. tip Recaps are the place to use a plan rather than a key. A meeting transcript is long, so it is the single most expensive thing you will send anywhere, and through a CLI bridge it uses no credit. ## Question 4: which language do you speak Language and engine are not independent. Pick the language first, then the engine that actually covers it. | Engine | Languages | Pick it when | | ------------------------ | ------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | | Whisper | Auto-detect plus 99 languages, the broadest by far. | You need a language outside Europe, or you switch languages often. | | Whisper Distil (English) | English only. | You only ever dictate in English and want roughly six times the speed of Large v3. | | Parakeet | European languages. It takes no language setting at all, so the Language box has no effect on it. | You speak a European language and want the fastest option on a Mac. | | Qwen3-ASR | Multilingual, but you should name the language rather than leave it on Auto. | You want the best result on conversational speech and acronyms. | warning Auto-detect is the default and works well on Whisper and Qwen. On Qwen, naming your language is still worth it: an unanchored long meeting has been measured coming back with its middle third in the wrong language. On Parakeet the setting does nothing, so forcing a non-European language there only degrades the result. ## Putting it together | If you are | Speech to text | Rewrite and recap | | -------------------------------------- | --------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------- | | On a Mac, and you want it to just work | Parakeet, the default. It runs on the Neural Engine, so a long meeting does not slow the rest of the Mac down. | Your Claude or ChatGPT plan through a CLI bridge, or a local model. | | On Windows with a 4 GB card or better | Whisper Base to start, Large v3 Turbo if accuracy is short. Qwen3-ASR 1.7B if your audio is messy conversation. | Cloud with your own key, or a local model if the card has room left. | | On Windows with integrated graphics | Cloud. Groq's Whisper Large v3 Turbo is the shipped default and it is fast. | Cloud, same key. | | Handling anything confidential | Local, any engine. Nothing leaves. | Local LLM, or off entirely. A cloud recap sends the whole transcript. | | Watching what you spend | Local. It costs nothing per minute. | A plan you already pay for, through a CLI bridge. | | Dictating only in English | Whisper Distil Large v3.5\. Same family, far quicker. | Whatever the rows above suggest. | ## Where to change it | What | Windows and macOS | Linux | | ----------------------- | ------------------------------------ | --------------------------------- | | Local or cloud | Settings → Voice input → Mode | Not exposed | | Local model | Settings → Voice input → Local model | Not exposed | | Cloud speech provider | Settings → Voice input → Provider | Settings → General → STT provider | | Language | Settings → Voice input → Language | Settings → General → Language | | Rewrite and recap model | Settings → Output | Settings → Output | | API keys | Settings → Providers & keys | Settings → General | info Linux is behind on this. The AppImage ships Whisper on the CPU only, with no Parakeet, no Qwen and no local model picker, so most of the local choices above are Windows and macOS today. tip Changing your mind is cheap. Models stay on disk until you remove them, and if you pick an engine whose files are missing, Dimmy falls back to Whisper and tells you it did rather than failing silently. Next: [how local mode works](https://dimmy.app/help/local-mode/), [the local model list](https://dimmy.app/help/whisper-models/), [the cloud providers](https://dimmy.app/help/cloud-providers/), [the on-device LLM](https://dimmy.app/help/local-llm/). Was this helpful? ## In this category [ Providers & ModelsLocal mode: transcription on your own machineFour speech engines that run on your device. No API key, no upload, no per-minute cost. Works on a plane.](https://dimmy.app/help/local-mode/) [ Providers & ModelsCloud providers comparedGroq is fastest, OpenAI is the gold standard, Deepgram leads on streaming. All optional. Bring your own key.](https://dimmy.app/help/cloud-providers/) [ Providers & ModelsAdd your API keyOpen Settings → Providers & keys, find the provider card, paste the key. It is encrypted on your machine. You can update keys at any time.](https://dimmy.app/help/api-keys/) [ Providers & ModelsChoose a local modelThe whole on-device catalogue: nine Whisper sizes, Parakeet TDT v3, four Qwen3-ASR variants and Whistle. What each one weighs and what it runs on.](https://dimmy.app/help/whisper-models/) [ Providers & ModelsOptional on-device LLMRun the LLM enhancement pass locally via llama.cpp. No cloud, no key, no per-token cost. Pick a model to match your GPU memory.](https://dimmy.app/help/local-llm/) Last updated: 2026-05-22 On this page 1. [01 Three choices, not one](#three-choices-not-one) 2. [02 Question 1: what machine do you have](#question-1-what-machine-do-you-have) 3. [03 Question 2: what must not leave your computer](#question-2-what-must-not-leave-your-computer) 4. [04 Question 3: which subscriptions do you already pay for](#question-3-which-subscriptions-do-you-already-pay-for) 5. [05 Question 4: which language do you speak](#question-4-which-language-do-you-speak) 6. [06 Putting it together](#putting-it-together) 7. [07 Where to change it](#where-to-change-it)