Hardware and platform requirements

Qualified platform

The current native package targets Windows x64. The tested machine has an NVIDIA GeForce RTX 4070 SUPER with roughly 12 GiB of VRAM and roughly 80 GiB of system RAM. That is a qualification machine, not a minimum specification or a promise that every 12 GiB card can run the same workload.

Linux, Windows ARM, Apple Silicon and MLX are roadmap targets. A compatible processor architecture alone is insufficient: the runtime binaries, drivers, model encodings, packaging, ownership and full generation flow must all be qualified for that platform.

Capacity is a resource tuple

Resource	Why it matters
Free GPU memory	Weights selected for offload, attention state and runtime buffers
Free system RAM	Remaining weights, host state, mappings and application overhead
CPU and memory bandwidth	Host-resident computation can dominate decode time
PCIe link	Device/host transfers can become a bottleneck
Storage and free disk	Model files, runtime components and temporary installation space
Driver and engine build	Required kernels and tensor encodings must be supported

A model can pass a coarse memory estimate and still fail because the engine lacks a kernel or needs a larger workspace. Conversely, a file larger than VRAM can sometimes run through host memory and partial offload, with a substantial performance tradeoff.

Backend choices

The current native controls expose CUDA and CPU. CUDA admission requires a detected NVIDIA GPU. The bundled CUDA distribution is selected for the tested hardware family; do not assume support for every GPU architecture from the presence of a CUDA label. CPU mode sets GPU placement to zero in the app’s load flow.

Keep the operating system responsive. RegesARC’s estimate reserves GPU and host headroom, but another application’s memory use can change after the snapshot. Close your own competing inference processes deliberately before measuring; RegesARC must not terminate unrelated services for you.

A sensible first run

Start with a complete supported GGUF, a modest context and explicit placement. Inspect System for real free memory before loading. If the model is larger than available resources, choose a smaller artifact deliberately or wait for a qualified execution strategy. Do not rely on SSD paging as an admitted substitute for RAM. See load planning and MoE architecture before interpreting a large parameter count as an active-memory requirement.