Gemma 4 26B-A4B inference in ~2 GB of RAM on any M-series MacBook
Swift apple-silicongemmagemma4gemma4-26b-a4bgpgpullmllm-inferencelocal-aimacosmetalon-device-aion-device-llmswift
import Metal
public struct VisionImagePlan {
public let metadata: VisionImageMetadata
public let geometry: Gemma4ImageGeometry
package let opened: OpenedVisionImage
package let started: ContinuousClock.Instant
}
public struct VisionPixelBuffer {
public let patchesBF16: MTLBuffer
public let positionsInt32x2: MTLBuffer
public let metadata: VisionImageMetadata
public let geometry: Gemma4ImageGeometry
public let wallNanoseconds: UInt64
public let allocatedBytes: Int
}