thermal yap · Windows · CUDA-only

Pick your card. Melt local.

Voxmelt runs two AI models on your GPU, so the GPU is not optional - it's the whole point. Any NVIDIA CUDA card from Ampere on up earns its keep here - GeForce RTX, workstation RTX A-series / Ada, or a data-center A100 / H100. Confirm the card you own and the download unlocks. AMD, Intel, and Mac sit this one out.

$ nvidia-smi --query-gpu=name,memory.total --format=csv
NVIDIA GeForce RTX 3080 Ti, 12288 MiB
NVIDIA CUDA GPU requiredCUDA-only · no AMD · no Intel · no Mac
Windows 10/11 · signed installer · ~12 MB on disk · 0 cloud calls
thermal budgeting

Box copy lies. Do the VRAM math.

Match your card against a cleanup model and watch the orchestrator decide: both coexist, auto-evict Whisper, or pick something smaller. Hardware honesty over marketing vapor.

compatibility check

Is your GPU ready to cook?

NVIDIA only · Win 10/11
Whisper 3 GB
4B 2.6 GB
0 GB6 GB12 GB
BOTH MODELS COEXIST · OVERCLOCKED W
Whisper stays warm. Zero handoff latency between recording and AI cleanup. Tensor-toasted throughput.
12 GB total VRAM3 GB Whisper2.6 GB Gemma 3 · 4B6.4 GB headroom
System requirements

What your rig needs

ComponentMinimumRecommended
GPUNVIDIA Ampere or newer · 8 GB VRAM (GeForce, workstation, or data-center)RTX 40/50-series or RTX Ada/A-series · 12+ GB VRAM
OSWindows 10 (64-bit)Windows 11
NVIDIA driver525.0+570.0+ (req. for Blackwell)
LLM runtimeOllama on 127.0.0.1:11434Ollama, latest
Disk~12 MB app + 3 GB model cacheSSD
NetworkFirst launch onlyNone after setup
All releases on GitHub