system requirements
What yapyap needs on macOS, Windows and Linux, and what each class of machine gets out of it.
yapyap runs every model on your machine. The models decide what you need, not the app. Every machine can record. How much RAM and which GPU you have decide how good the models are and how long you wait.
the short answer
| machines | what you get | |
|---|---|---|
| great | Apple Silicon with 16 GB or more. A PC with 32 GB and a dedicated GPU. | Transcripts in minutes. A 7B or 14B language model writes the summaries. |
| good | Apple Silicon with 8 GB. A PC with 16 GB and a recent integrated GPU. | Transcripts in about 5 to 20 minutes per hour of audio. A small language model writes shorter, simpler summaries. |
| slow | A PC with no usable GPU. | Transcripts take 10 to 20 minutes per hour of audio in 25 European languages. Other languages can take 2 to 4 hours per hour of audio. |
Recording is safe on all of them. Audio capture always comes before the AI work.
operating system
| macOS | Windows | Linux | |
|---|---|---|---|
| OS | macOS 11 (Big Sur) or newer | 64-bit Windows 10 or newer | Ubuntu 24.04+, Debian 13+, Fedora 40+, or any distribution with glibc 2.39+ |
| CPU | Apple Silicon (M1 or later). Intel Macs are not supported. | any 64-bit Intel or AMD processor | any 64-bit Intel or AMD processor |
| GPU | built in. Transcription uses the Apple GPU through Metal. | optional, but it changes a lot (see below) | optional, but it changes a lot (see below) |
The Windows installer brings the WebView2 runtime, the Microsoft C++ runtime and the Vulkan loader. You do not install anything separately.
On Linux, the deb and rpm packages need webkit2gtk 4.1 and pull it in for
you. The AppImage bundles it.
memory: what each amount of RAM gets you
The setup wizard reads your RAM and suggests a preset. It keeps 30% of your RAM (at least 4 GB) free for the system, your browser and your meeting app. The rest goes to the models.
| your RAM | suggested preset | models in memory at peak | language model |
|---|---|---|---|
| 8 GB | featherweight | about 3.5 GB | Qwen 2.5, 1.5B |
| 12 GB | balanced | about 5 GB | Qwen 2.5, 3B |
| 16 to 24 GB | performance | about 9 GB | Qwen 2.5, 7B |
| 32 GB or more | studio | about 16 GB | Qwen 2.5, 14B |
The language model is part of every preset. It writes the titles, the summaries and the lens results. On 8 GB it is the smallest model we ship. That model works, but its summaries are short and miss detail. For summaries you can rely on, plan for 16 GB.
The figures above count the model weights only. The app window and the working memory of each model come on top. On 8 GB, close what you do not need while yapyap processes a long recording.
A machine with less than 8 GB gets the essential preset. It turns off punctuation, names and sentiment, and uses the smallest models. You can change every preset later in settings.
GPU: the biggest difference on Windows and Linux
On a Mac, every Apple Silicon chip has a GPU that yapyap uses. On a PC, the GPU decides how long you wait for a transcript.
| PC | transcribing one hour of audio |
|---|---|
| dedicated GPU with Vulkan (Nvidia, AMD) | about 5 to 10 minutes |
| integrated GPU with Vulkan (Intel Iris Xe, AMD 780M and similar) | about 20 minutes |
| no usable GPU, app language in Parakeet's 25 | about 10 to 20 minutes |
| no usable GPU, any other language | about 2 to 4 hours |
These are estimates from the app's own speed model, not benchmarks of your machine. Clock speed, memory speed and other open apps move them.
Why the last row is so slow. On Intel and AMD processors, the whisper transcription engine runs without the processor's vector instructions. That keeps yapyap working on old processors and virtual machines, but it makes whisper 4 to 8 times slower on the CPU. A GPU removes that cost.
Parakeet is the way around it. Parakeet is a second transcription engine that runs fast on the CPU. It covers 25 European languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. The wizard picks Parakeet for the three lighter presets (essential, featherweight, balanced) when the app language is one of these. You can also pick it yourself in settings. For other languages, whisper is the only choice, and a GPU is what makes it quick.
The language model does not use an integrated GPU. On a PC without a dedicated GPU, summaries run on the CPU at about 10 words per second. A summary takes a minute or two instead of a few seconds.
disk
| preset | models on disk |
|---|---|
| essential | about 2 GB |
| featherweight | about 3.5 GB |
| balanced | about 5.5 GB |
| performance | about 9 GB |
| studio | about 15 GB |
Your recordings come on top of this. An SSD loads the models faster and keeps downloads from slowing a recording. The app shows every download before it starts. Nothing goes onto your disk without your consent.
tips for a smaller machine
- Let the first-run model downloads finish before your first real recording.
- Leave live transcription off. It is off by default.
- Expect the transcript after you stop recording, not during it.
- On a PC without a GPU, pick Parakeet in settings if you record in one of its 25 languages.
checking what you have
-
Windows: press Ctrl+Shift+Esc and open the Performance tab. It shows the CPU, the core count, the RAM and the GPU. Or in PowerShell:
Get-CimInstance Win32_Processor | Select-Object Name, NumberOfCores [math]::Round((Get-CimInstance Win32_ComputerSystem).TotalPhysicalMemory/1GB,1) Get-CimInstance Win32_VideoController | Select-Object Name -
macOS: open the Apple menu and choose About This Mac.
-
Linux:
lscpu,free -handvulkaninfo --summary(if installed) show the CPU, the RAM and whether a Vulkan driver is present.
a note on older machines
Since v0.6.8, the Windows and Linux builds run on any 64-bit x86 processor. That includes older CPUs and virtual machines without modern vector instructions (AVX), and machines with no Vulkan driver at all. On such machines, transcription is slow (see the GPU table above).
Versions before v0.6.8 crashed without an error at launch or during onboarding on that hardware. If that happens to you, install the latest version.