System requirements
Compare Windows, RAM, storage, CPU and GPU guidance for local Whisper and Qwen models with an interactive VRAM budget chart.
Last updated
Baseline requirements
Local Mode runs Whisper through whisper.cpp. Speed depends on the model, memory, driver and CPU or GPU. Cloud needs an internet connection but uses fewer PC resources.
| Component | Baseline |
|---|---|
| Operating system | Windows 10 version 1809 or later (64-bit) or Windows 11 (64-bit) |
| Storage | SSD with free space for the application, working files and installed models |
| Processor | Recent 4-core CPU minimum; larger models benefit from more cores |
| Graphics | Optional GPU acceleration; CPU processing remains available |
| Graphics driver | Current manufacturer driver recommended for GPU processing |
| Internet | Needed for installation downloads, model downloads, licenses, updates and Cloud |
Choose by model
Use the selector to compare the six models included with WhisperARC. The installer starts with Large v3 Turbo q5. Large v3 Turbo q8 is a sensible next step if your computer has a stronger dedicated GPU.
These figures leave room for Windows and long recordings. Treat them as a starting point rather than a performance promise.
Choose a model then check the hardware.
Installer default for balanced quality, speed and storage.
- Accuracy
- Very good
- Speed
- Very fast
- Model file
- 547 MB
- Memory
- 16 GB RAM
- Processor
- Recent 4-core / 8-thread CPU or better
- Graphics
- Optional. A dedicated Vulkan GPU with 4 GB+ VRAM is recommended. CPU fallback is slower.
- Storage
- Keep 4 GB free for WhisperARC, shared components and working space.
- This is the model currently preselected by the WhisperARC installer.
- Whisper large-v3-turbo keeps the large-v3 family but reduces decoder layers from 32 to 4. This improves speed with a small quality tradeoff.
- Q5 quantization reduces disk and memory pressure compared with heavier variants.
Model comparison
| Model | File | Memory | Graphics recommendation | Best fit |
|---|---|---|---|---|
| Base q8 | 78 MB | 8 GB RAM | Optional. WhisperARC uses Vulkan when available and otherwise runs on the CPU. | Older or modest PCs and short dictations. |
| Small q8 | 252 MB | 8 GB RAM | Optional. A Vulkan GPU improves speed; CPU fallback is automatic. | Everyday dictation on light laptops and entry desktop PCs. |
| Large v3 Turbo q5 | 547 MB | 16 GB RAM | Optional. A dedicated Vulkan GPU with 4 GB+ VRAM is recommended. CPU fallback is slower. | Installer default for balanced quality, speed and storage. |
| Large v3 Turbo q8 | 834 MB | 16 GB RAM | Optional; a dedicated NVIDIA, AMD or Intel Vulkan GPU with 6 GB+ VRAM is recommended. CPU fallback is slower. | Higher precision on strong consumer PCs with a dedicated GPU. |
| Large v3 q5 | 1.01 GB | 16 GB RAM minimum; 32 GB preferred for long files | Optional. A dedicated Vulkan GPU with 8 GB+ VRAM is recommended. CPU fallback can be substantially slower. | Quality-sensitive work on powerful desktops and recent gaming laptops. |
| Large v3 Turbo F16 | 1.51 GB | 16 GB RAM minimum; 24 GB preferred for long files | Optional. A dedicated Vulkan GPU with 8 GB+ VRAM is recommended. CPU fallback is slower. | Highest Turbo quality on strong desktops and recent gaming laptops. |
Qwen local requirements and VRAM
Qwen handles local translation and Assistant modes. Select the exact model used in WhisperARC below then compare its advised VRAM budget with the memory available on your GPU.
RAM, CPU and VRAM values are conservative WhisperARC planning estimates for short text processing. They are not official Qwen minimums or benchmarked memory peaks. They allow headroom beyond the model file for inference buffers and context; CPU generation can still be slow.
| Model | Model file / free SSD for download | System RAM advised | CPU advised | Qwen VRAM budget advised |
|---|---|---|---|---|
| 3.16 GiB / 6.9 GiB free | 16 GB | Recent 6-core / 12-thread CPU | 8 GB; GPU optional | |
| 1.81 GiB / 4.2 GiB free | 16 GB | Recent 4-core / 8-thread CPU | 6 GB; GPU optional | |
| 0.96 GiB / 2.5 GiB free | 8 GB | Recent 4-core CPU | 4 GB; GPU optional | |
| 6.14 GiB / 12.8 GiB free | 24 GB | Recent 8-core / 16-thread CPU | 12 GB; GPU optional |
Allow for Windows, Whisper and other GPU applications before choosing the available amount.
Recommended budget: 8 GB. Available ceiling: 8 GB. Matches the advised budget with no extra margin.
The filled bar is a recommended Qwen budget; the vertical line is your available ceiling. This estimates headroom, not measured usage or a guaranteed peak. A smaller GPU may work with partial CPU processing.
Sources and estimation method
The Qwen model cards describe the original model families, not WhisperARC's quantized file sizes or Windows hardware minimums. File sizes and quantizations above come from the verified WhisperARC model catalogue; GiB means 1,073,741,824 bytes. Free SSD space follows the app's download check: twice the file size plus 0.5 GiB, rounded up. Allow separate space for WhisperARC, its runtimes, Whisper models and recordings.
WhisperARC runs Qwen with an 8,192-token context, one request at a time and one Qwen model loaded. Ollama documents that longer contexts and concurrent requests increase memory needs. Its model can remain loaded for five minutes. Whisper uses a separate runtime so their memory use can overlap.
GPU compatibility depends on Ollama's bundled backend and your driver separately from Whisper's Vulkan support. WhisperARC enables Ollama's Vulkan path alongside available CUDA or ROCm support. Check the official Windows GPU list below and test on your own PC; GPU detection alone does not guarantee full acceleration.
- Qwen: official 4B model card and architecture
- Qwen: official 2B model card and architecture
- Qwen: official 0.8B model card and architecture
- Qwen: official 9B model card and architecture
- Ollama: context length and memory requirements
- Ollama: model residency, concurrency and GPU memory
- Ollama: Windows GPU and driver compatibility
Real-world performance
Laptop power limits, long files, translation, other graphics applications and CPU fallback can change processing time. Test a representative recording on the target computer before choosing hardware for a production workflow.