← Back to Help center
Providers & Models

Choose a local model

The whole on-device catalogue: nine Whisper sizes, Parakeet TDT v3, four Qwen3-ASR variants. What each one weighs and what it runs on.

Open Settings → Voice input → Local model. On macOS the picker is grouped into Whisper, Parakeet and Qwen3-ASR, and every row names the processor it will use. On Windows it is one flat list with a VRAM dot per row. Linux does not expose a local model picker at all: the AppImage runs Whisper on the CPU.

The local model picker in Settings, with sizes and download state per row

Three engines share that one picker. If you are not sure which family you want, start from Which model should you use?. What follows is the reference: what exists, what it weighs, what it is good at.

Whisper

The broadest option: 99 languages, nine sizes, the same list on every platform.

ModelFileSizeGood for
Tinyggml-tiny-q8_0.bin42 MBFastest, lower accuracy
Baseggml-base-q8_0.bin78 MBGood balance of speed and accuracy. The Whisper default
Smallggml-small-q5_1.bin181 MBHigh accuracy, slower
Mediumggml-medium-q5_0.bin514 MBVery high accuracy, wants 2 GB or more of RAM
Large-v3-Turbo Q5ggml-large-v3-turbo-q5_0.bin574 MBFast and accurate, all languages
Large-v3-Turbo Q8ggml-large-v3-turbo-q8_0.bin874 MBBest turbo quality, all languages
Large-v3 Q5ggml-large-v3-q5_0.bin1.1 GBMaximum accuracy, all languages, slow
Distil-Large-v3.5 Q8ggml-distil-large-v3.5-q8_0.bin818 MBEnglish only, around 6x faster than Large-v3
Distil-Large-v3.5 Q5ggml-distil-large-v3.5-q5_0.bin538 MBEnglish only, compact

The q5_0, q5_1 and q8_0 in the file names are quantisation levels. A lower number means a smaller file and slightly coarser weights. Dimmy only ships quantised builds, so there is no full-precision row to compare against.

Parakeet TDT v3

One row in the picker, two very different builds underneath. On macOS it is a CoreML bundle of about 466 MB that runs on the Neural Engine. On Windows it is an ONNX bundle of about 2.5 GB that runs on the CPU. Same model, same quality, wildly different download.

  • Fast, and strong on European languages.
  • It takes no language argument at all, so the Language setting has no effect on it. See Pick the dictation language.
  • On Windows, anything longer than two minutes is split and stitched back together automatically, because the model itself breaks somewhere past five minutes of audio in one go.

Qwen3-ASR

The newest engine, and the one that measured best on messy conversational speech and on acronyms. It needs its language named rather than left on Auto.

VariantSizeRuns onGood for
Qwen3-ASR 0.6B971 MBGPU or CPUTwice as fast, weaker on acronyms
Qwen3-ASR 1.7B2.4 GBGPU or CPUBest accuracy on conversational speech. The Qwen default
Qwen3-ASR 0.6B · Neural Engine (int8)2.9 GBNeural EngineLeaves the GPU to the rest of the Mac. Needs macOS 15
Qwen3-ASR 0.6B · Neural Engine (f32)4.0 GBNeural EngineFull precision on the Neural Engine. Needs macOS 15

Each GPU variant is a pair of files, the model plus an audio projector, and both have to be on disk before the row counts as downloaded. The two Neural Engine variants only appear on Apple Silicon running macOS 15 or newer; on any other machine they are not listed rather than offered and then failing.

What the rows tell you

  • A green check means the model is already on disk. Rows you have not downloaded show their size instead.
  • On macOS each row also names where it will run. A Whisper row says Neural Engine only when that model's own Core ML encoder is on disk, and GPU otherwise.
  • On Windows each row can carry a coloured dot read from your graphics card: green means it fits comfortably, amber means it fits but may fail once Whisper is also resident, red means it is too large. No dot means Dimmy could not read your VRAM and will not guess.
  • Nothing is ever hidden or blocked because of your hardware. The dots are advice, not a gate.

Download flow

  1. Pick a model from the list.
  2. Click Download. A progress bar shows the current MB of the total, and an interrupted download resumes where it stopped rather than starting over.
  3. When it finishes the model becomes active automatically.
  4. Earlier models stay on disk.

The Neural Engine row on macOS

Whisper's encoder can move off the GPU and onto the Neural Engine, which is the difference between a long meeting slowing Dimmy down and a long meeting slowing the whole Mac down. Dimmy offers the encoder as a separate download for the selected model, around 1.2 GB for the large ones. The first transcription afterwards takes a few minutes while macOS compiles the model for your specific machine, once. The Distil models have no Core ML encoder, so no offer appears for them.

Was this helpful?

In this category

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.
Providers & Models
Local mode: transcription on your own machine
Three speech engines that run on your device. No API key, no upload, no per-minute cost. Works on a plane.
Providers & Models
Cloud providers compared
Groq is fastest, OpenAI is the gold standard, Deepgram leads on streaming. All optional. Bring your own key.
Providers & Models
Add your API key
Open Settings → Providers & keys, find the provider card, paste the key. It is encrypted on your machine. You can update keys at any time.
Providers & Models
Optional on-device LLM
Run the LLM enhancement pass locally via llama.cpp. No cloud, no key, no per-token cost. Pick a model to match your GPU memory.
Last updated: 2026-05-22