Some things you say out loud should stay in the room. A draft to a client. Notes about a patient. The half-formed idea you are not ready to show anyone yet.
Until now, Meander sent your voice to the cloud to turn it into text. That is fast, and it is what most people want. But we kept hearing the same question from Linux users, privacy folks, and anyone with a decent graphics card sitting idle: can it run on my own machine?
It can now. Meander's desktop app can run both halves of voice typing on your computer: the speech recognition that turns your voice into text, and the AI cleanup that fixes the punctuation and filler words. Your audio never leaves the machine.
What runs locally
Voice typing in Meander has two steps, and you can move each one on its own:
- Speech recognition runs on whisper.cpp. You pick from the Whisper family, from Tiny (78 MB) up to Large v3, plus Breeze ASR 25 for Taiwanese Mandarin and Chinese-English mixing.
- AI cleanup runs on llama.cpp. You pick a small language model: Gemma, Qwen, Llama, Mistral and others, from under 1 GB up to around 20 GB.
Both use your GPU through Vulkan when you have one, and fall back to the CPU when you don't. Once you are signed in, dictation keeps working with no internet connection.
Set it up in 3 steps
1. Download the models. Open Models in Meander. Every model shows its size, a speed and accuracy rating, and a download button. For a first setup we recommend Whisper Large v3 Turbo q5 (574 MB) for speech and Gemma 4 E2B (3.5 GB) for cleanup.
2. Point a mode at them. Open Modes, create a new mode (or edit one you already use), and set STT model to Local with Turbo q5.
Then set LLM model to Local with Gemma 4 E2B.
3. Give it a hotkey and talk. Assign a key to the mode, press it, and speak. The text lands wherever your cursor is, same as before. The first time, Meander downloads its local engine, which takes a moment.
Why those two models? We timed every option on real dictation before picking. Turbo q5 was both faster and more accurate than the full Large v3, and Gemma 4 E2B was the most faithful cleanup model, fixing the text without rewriting it. The full results, including how long you wait on CPU versus GPU, are in our local vs cloud benchmark.
Mix and match
Local and cloud are not a single switch. Each mode picks its own speech and cleanup source, so you can build the setup that fits the moment:
- Everything local for private work. Nothing leaves your computer.
- Local speech, cloud cleanup as a middle ground. Your audio stays on your machine, and only the text goes out.
- Everything cloud on a laptop without a GPU, when speed matters most.
Switching is as quick as switching modes. Keep a private mode and a fast mode side by side and use whichever the moment calls for.
Already running Ollama or LM Studio?
You don't have to use our model library. Meander can point any mode at your own OpenAI-compatible endpoint, such as Ollama, LM Studio, or a server on your network. The API key is optional for servers that don't use one.
One GPU for the whole house
If one computer has the graphics card and the others don't, Meander can share its downloaded models over your local network. Turn on Share on LAN on the machine with the GPU, and the laptops in the house can use its models as a custom endpoint. Access is protected by a token, and the setting applies after a restart.
Who can use it
Local and custom models are part of Meander Pro, and they are included in the free 2-week Pro trial. Share on LAN is part of Ultimate. Local mode runs on the Meander desktop app for Linux and Windows.
Local runs don't use your cloud quota, so long recordings and heavy days cost nothing extra.
If you want to see which local speech models are worth knowing about beyond Whisper, we wrote a roundup of the best local voice-to-text models.
Ready to try it? Download Meander. On Linux, here is how to set it up on Wayland.