diff --git a/discover/cuda_common.go b/discover/cuda_common.go
index 3c7a92114..a2c43420e 100644
--- a/discover/cuda_common.go
+++ b/discover/cuda_common.go
@@ -16,7 +16,7 @@ import (
 // Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
 var CudaTegra string = os.Getenv("JETSON_JETPACK")
 
-func cudaVariant(gpuInfo CudaGPUInfo) string {
+func cudaVariant(gpuInfos []CudaGPUInfo) string {
 	if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
 		if CudaTegra != "" {
 			ver := strings.Split(CudaTegra, ".")
@@ -45,20 +45,19 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 		}
 	}
 
-	// Check GPU compute capability FIRST
-	isOldGPU := gpuInfo.computeMajor < 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor < 5)
-	if isOldGPU {
-		// GPU is Pascal or older (CC <= 7.4) - use CUDA v12 (supports CC 6.1)
-		return "v12"
+	// Check GPU compute capability FIRST, lowest common denominator if multi-gpu
+	for _, gpuInfo := range gpuInfos {
+		if gpuInfo.computeMajor < 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor < 5) {
+			// GPU is Pascal or older (CC <= 7.4) - use CUDA v12 (supports CC 6.1)
+			return "v12"
+		}
 	}
 
 	// GPU is Turing or newer (CC >= 7.5) - can use newer CUDA
-	if gpuInfo.DriverMajor < 13 {
+	if len(gpuInfos) > 0 && gpuInfos[0].DriverMajor < 13 {
 		// The detected driver is older than 580 (Aug 2025)
 		// Warn if their CC is compatible with v13 and they should upgrade their driver to get better performance
-		if !isOldGPU {
-			slog.Warn("old CUDA driver detected - please upgrade to a newer driver for best performance", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
-		}
+		slog.Warn("old CUDA driver detected - please upgrade to a newer driver for best performance", "version", fmt.Sprintf("%d.%d", gpuInfos[0].DriverMajor, gpuInfos[0].DriverMinor))
 		return "v12"
 	}
 	return "v13"
diff --git a/discover/gpu.go b/discover/gpu.go
index b09626118..a39bc7c3d 100644
--- a/discover/gpu.go
+++ b/discover/gpu.go
@@ -284,18 +284,8 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.MinimumMemory = cudaMinimumMemory
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
-				variant := cudaVariant(gpuInfo)
 
-				// Start with our bundled libraries
-				if variant != "" {
-					variantPath := filepath.Join(LibOllamaPath, "cuda_"+variant)
-					if _, err := os.Stat(variantPath); err == nil {
-						// Put the variant directory first in the search path to avoid runtime linking to the wrong library
-						gpuInfo.DependencyPath = append([]string{variantPath}, gpuInfo.DependencyPath...)
-					}
-				}
 				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-				gpuInfo.Variant = variant
 
 				if int(memInfo.major) < cudaComputeMajorMin || (int(memInfo.major) == cudaComputeMajorMin && int(memInfo.minor) < cudaComputeMinorMin) {
 					unsupportedGPUs = append(unsupportedGPUs,
@@ -333,6 +323,24 @@ func GetGPUInfo() GpuInfoList {
 				// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
 				cudaGPUs = append(cudaGPUs, gpuInfo)
 			}
+			// Second pass on NVIDIA GPUs to set lowest common denominator variant and DependencyPaths
+			variant := cudaVariant(cudaGPUs)
+			var variantPath string
+			// Start with our bundled libraries
+			if variant != "" {
+				variantPath = filepath.Join(LibOllamaPath, "cuda_"+variant)
+				if _, err := os.Stat(variantPath); err != nil {
+					variantPath = ""
+				}
+			}
+
+			for i := range cudaGPUs {
+				cudaGPUs[i].Variant = variant
+				if variantPath != "" {
+					// Put the variant directory first in the search path to avoid runtime linking to the wrong library
+					cudaGPUs[i].DependencyPath = append([]string{variantPath}, cudaGPUs[i].DependencyPath...)
+				}
+			}
 		}
 
 		// Intel