APERTURESandhu Consulting Group
Version 0.4.6 Release guide

Open-source local inference / 0.4.6

APERTURE

Configure local inference around your chosen model.

Supply a model link or a local file. Aperture inspects your hardware with permission, explains a CPU/GPU configuration, and opens a local session through a supported runtime.

Your selected artifact and explicit context remain unchanged. Save the configuration, return to a local chat, or run a bounded experiment on the same machine.

Get the install command →See demonstrated support

From a chosen model to an explained method

  1. 01

    Inspect the machine

    Approve a read-only view of hardware and available resources.

  2. 02

    Choose the artifact

    Paste a model link or local path. Select the exact representation.

  3. 03

    Review the method

    See the candidate backend, memory budgets, context, and limits.

  4. 04

    Start a session

    Approve acquisition and runtime installation when required, then run.

This describes the terminal workflow. The website does not scan your computer.

Using the release

A model-first workflow with explicit choices

01 / INPUT

Use a link or an existing model

Inspect supported GGUF files, split sets, repository links, and local folders. Existing files remain in place.

Model sources →
02 / PLACEMENT

Understand where the work runs

For compatible GGUF models, one GPU can execute some layers while the CPU handles the rest. Memory pools stay distinct.

Memory and placement →
03 / CONTROL

Keep permissions separate

A scan does not approve downloads, installation, inference, or experiments. Results are not uploaded automatically.

Privacy and local state →

Observed support

Check the execution path and its requirements

Native records cover Windows CPU, selected NVIDIA and Intel graphics paths, and Linux CPU. Version 0.4.1 includes a recorded RTX 3090 CUDA chat after the installed-library discovery correction.

Partial offload above available VRAM has been observed on a 4060. Successful execution above its physical capacity remains unverified. NPU inventory is not NPU execution.

Read the support and evidence matrix →
Managed GGUF runtime
node-llama-cpp 3.20.0
Guided concurrency
One sequence
Default context
4,096 tokens; explicit values preserved
External harness connections
Not implemented in this release

Documentation

Installation, methods, and recovery in one place

Get startedInstallation, permissions, the first session, and returning later.Model sourcesLocal files, repository links, numbered shards, and access.Memory and placementPhysical capacity, available headroom, and CPU/GPU split execution.Command referenceEvery released command, flag, and independent permission.TroubleshootingRecover from installation, download, backend, and model errors.Privacy and local dataWhat is read, retained, transmitted, and kept under your control.Bounded experimentsRun the optional comparison and interpret its local evidence.Verified supportImplemented paths, native observations, and limits of the evidence.ReleasesVersioned changes, pending work, and maintenance guidance.MaintenanceBuild, verify, and publish the release-bound documentation.