Skip to content

Desktop packages

Fabrizio Degni edited this page Nov 15, 2026 · 2 revisions

Desktop packages: architecture and execution backend

Package identity describes software requirements, not a particular computer model. Use NodeLift-<component>-<version>-<os>-<architecture>[-<backend>]-<delivery>.zip for new worker distributions. Existing published filenames remain valid.

x64, AMD64 and x86-64 identify the same 64-bit architecture on both Intel and AMD processors. x86 here means 32-bit Windows. arm64 means Windows AArch64 for these packages; it does not imply compatibility with Android, iOS or Linux.

Choose a package

Package Included Requirements and status
Node 0.7.1-preview.1, Windows x86 / x64 / ARM64 Native launcher, Python, NumPy/CPU BLAS, inventory, enrollment and bounded CPU benchmarks Public qualification preview. No system Python required. Native Windows startup and physical performance remain untested. No LLM executor or GPU/NPU execution bundled.
NodeLift-Worker-0.6.2-windows-x64-cpu-cuda-portable.zip Python, CPU/CUDA runtime, dependencies and a 491 MB control GGUF Private qualification kit. Extract the whole folder and open Open-NodeLift.vbs. Requires saved device enrollment. CUDA requires a compatible NVIDIA GPU/driver; CPU control is a separate profile.
NodeLift-Worker-0.6.1-windows-x64-cpu-cuda-update.zip Application update and RPC executable Private qualification update. Reuses Python, model and complete CPU/CUDA dependencies from the legacy installation. It cannot replace the portable package on a new computer.

The portable archive is approximately 1.2 GB; the application update is approximately 0.1 MB because the runtime and model are already installed. This is a packaging difference, not a performance or hardware-tier distinction. Run only one worker.

The public generic preview supports first pairing by invitation. The two advanced qualification kits retain test-network broker configuration and require an existing association. They are served through the private test channel, not the public release assets. No permanent LAN address is part of a public package identity.

Backend selection and evidence

CPU and CUDA are execution backends. CUDA requires NVIDIA hardware and a usable driver; an AMD or Intel x64 CPU alone says nothing about CUDA availability. The portable worker includes CPU instruction variants selected by the native runtime. Detection, successful library loading, numerical qualification and measured speed are distinct states. Neither this rename nor cross-compilation proves acceleration.

These worker archives do not include a Vulkan backend. An included ONNX Runtime DirectML dependency supports a separate probe path; it does not establish a DirectML LLM executor. Metal, Vulkan, other GPU backends and NPU support require their own OS-specific build, supported operators and physical qualification before appearing as supported execution options. The Node preview currently executes CPU math only, even when inventory detects a GPU.

Compatibility

The previous NodeLift-P16-* download URLs remain available privately for existing transfers. They refer to the old bytes and hashes; new canonical archives have a standard folder name, updated interface labels and new checksums. Historical benchmark reports continue to name the measured computer so evidence stays attributable. Legacy script names and the quick update's existing C:\dev\NodeLift-P16-qualification path are compatibility details, not hardware requirements for the portable worker.

See generic node setup and validation, multi-device test setup and packaging verification.

Clone this wiki locally