Model and GPU guidance
Understand how GPU guidance changes with the Koeva transcription model you choose.
GPU fit by model
Koeva uses Vulkan acceleration when a compatible runtime and graphics device are available. Otherwise it switches automatically to CPU mode and uses half of the computer's logical CPU cores. Base and Small remain the best choices for lighter hardware; Turbo and Large benefit much more from a dedicated GPU.
| Model | File | Memory | Graphics recommendation | Best fit |
|---|---|---|---|---|
| Base q8 | 78 MB | 8 GB RAM | Optional. Vulkan acceleration is used when available; otherwise Koeva runs automatically on CPU. | Older or modest PCs, quick drafts, 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; dedicated Vulkan GPU with 4 GB+ VRAM recommended. CPU fallback is slower. | Balanced quality on mid-range machines. |
| Large v3 Turbo q8 | 834 MB | 16 GB RAM | Optional; dedicated NVIDIA, AMD, or Intel Vulkan GPU with 6 GB+ VRAM recommended. CPU fallback is slower. | Recommended default for strong consumer PCs. |
| Large v3 q5 | 1.01 GB | 16 GB RAM minimum; 32 GB preferred for long files | Optional; dedicated Vulkan GPU with 8 GB+ VRAM recommended. CPU fallback can be substantially slower. | Higher accuracy on powerful desktops and recent gaming laptops. |
| Large v3 | 2.88 GB | 32 GB RAM recommended | Optional; dedicated Vulkan GPU with 10 GB+ VRAM recommended. CPU fallback can be substantially slower. | Best quality on high-end desktops or workstation-class laptops. |
How the guidance was chosen
The guidance above uses four reference points: OpenAI's model-size and VRAM table for Whisper, OpenAI's large-v3-turbo model card, the whisper.cpp quantization notes, and the public GGML model repository. Quantized GGML models reduce disk and memory pressure, but exact runtime memory and speed still depend on the backend and computer.
Koeva has not benchmarked every GPU. Treat the VRAM guidance as a practical starting point, not as an exhaustive compatibility list.
Keep drivers current
A current, stable graphics driver is needed to use Vulkan acceleration reliably. If Vulkan cannot be initialized, Koeva keeps working through its automatic CPU fallback.
Buying a computer
When choosing a new Windows computer for Koeva, decide which model you expect to use most. For Base or Small, prioritize a recent CPU and enough RAM. For Turbo q8 or either Large model, prefer a dedicated NVIDIA GeForce, AMD Radeon, or Intel Arc GPU with the VRAM shown in the model table.