The engine, not the wrapper.

A purpose-built engine written in Rust owns model loading, memory management, the compute graph, and dispatch to GPU kernels. No garbage-collector pauses, tight control over memory so large models fit on consumer devices, and direct integration with hand-tuned kernels, for up to 60% faster inference than off-the-shelf runtimes.

in this build ·This web build keeps the engine's role (load, stream, token accounting) behind a stable interface, backed by WebGPU. The native app ships the real Rust engine.

Rather than lean on generic libraries, Aperture writes its own GPU kernels for Apple (Metal) and NVIDIA (CUDA). Fused operations (quantized matmul, RMSNorm, RoPE, and attention) cut memory traffic, and a flash-attention implementation avoids ever materializing the full attention matrix.

in this build ·Browsers can't ship Metal/CUDA. The web build uses WebGPU, which itself compiles down to Metal/Vulkan/D3D, for real on-device compute on the /local and /gateway pages.

Weights are compressed from 16-bit floats to ~4-bit integers with the Q4_K scheme: block-wise quantization with per-block scales for a strong quality-to-size tradeoff. That is a ~4x smaller memory footprint and far less memory bandwidth per token, which is the difference between a model fitting on a laptop or not.

in this build ·The web runtime loads already-quantized models, so quantization shows up as a property of the model you run (e.g. 'Q4_K · ~0.9 GB') rather than a step the web build performs itself.

On Apple Silicon the CPU and GPU share one pool of memory, so model weights are addressable by the GPU with no copy. Aperture tunes its engine and Metal kernels around this unified-memory architecture to squeeze maximum throughput out of the Mac in front of you.

in this build ·WebGPU maps to Metal on a Mac, so the browser path already benefits from Apple Silicon implicitly. The '60% faster' figure describes the native app and is presented here as such.