← Back to Help center
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.

In local mode the recording never leaves the machine. On first launch Dimmy downloads a model: around 78 MB for Whisper base-q8_0 on Windows, or the Parakeet bundle of about 466 MB on macOS, where Parakeet is the default engine. After that, transcription is free and works with no connection.

What runs locally

  • Speech to text, with three engines to choose between: whisper.cpp, Parakeet TDT v3 and Qwen3-ASR. See Choose a local model.
  • Filler removal (no model, just regex).
  • LLM rewriting, optional, if you enable the on-device model. See Optional on-device LLM.
  • Everything else: history, settings, key storage.

Where it runs on each platform

PlatformSpeech runs onEngines available
macOS, Apple SiliconMetal on the GPU, plus the Neural Engine for Parakeet, for Qwen's Neural Engine variants, and for Whisper once you download its Core ML encoder.Whisper, Parakeet, Qwen3-ASR
WindowsVulkan on the GPU, falling back to the CPU on its own when the driver reports no usable device.Whisper, Parakeet, Qwen3-ASR
Linux, AppImageCPU only. The AppImage stays CPU-only so it runs across distributions with different Vulkan loaders.Whisper

Performance

Measured on a 4 GB NVIDIA T600 over Vulkan, same 15-second chunk in the same session: Whisper Base 1.5 s, Whisper Small 1.7 s, Whisper Large-v3-Turbo Q8 5.9 s, Qwen3-ASR 1.7B 2.6 s. Parakeet is the odd one out because its cost tracks the audio you actually give it: measured on the CPU alone, warm latency was 337 to 547 ms.

image · placeholder
Local benchmark chart: latency by hardware and clip length

When to use cloud instead

  • You are on integrated graphics, where local models load but crawl.
  • You record long sessions and want them back in seconds without tying up the GPU.
  • You want realtime typing at the lowest possible latency. That path uses OpenAI Realtime or Deepgram when you have saved a key for either.
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
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
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.
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