## v1.2.011 — 2026-06-20
- fix: **On PCs with two GPUs — an integrated chip plus a dedicated graphics card — the OmniVoice high-quality voice and voice cloning no longer run on the weak integrated GPU.** On these systems the GPU voice engine could select the integrated GPU instead of the dedicated card, which made each line take roughly 50 seconds to generate and then time out — so pressing "Test voice" or saving a cloned voice appeared to do nothing at all. The app now detects the installed GPUs and pins the voice engine to the dedicated graphics card, restoring fast generation. Systems with a single GPU are unaffected.

- fix: **The local AI now receives the same multi-GPU handling on its Vulkan backend.** When the local model runs on the Vulkan backend — used on non-NVIDIA cards, or when the NVIDIA CUDA path is unavailable — it would by default spread the model across every detected GPU, including a weak integrated one, slowing generation. It now targets the dedicated graphics card only.