Free desktop beta · Windows and Linux x64

OutilsIA Local Cockpit

Find a local AI model that fits your PC. Test what actually runs before buying more hardware.

Public 0.1.1-beta · build 20260928000500. The app interface is currently in French. No macOS or ARM64 package in this release.

Check your PC

See detected CPU, RAM, GPU and runtime information. Compatibility suggestions are estimates, not measured speed.

Run a real test

Select a local model, review the operation and start a benchmark. Loading, generation and GPU placement are separate observations.

Keep the result

Generate a report, inspect the limits and decide what to try next. Buying nothing is a valid outcome.

Local Cockpit interface preview in French, generated from the September 13 source
Current-source interface preview in French. This browser rendering is not a native hardware test or a benchmark result.

Your first run

  1. Download and check the fileUse the matching SHA-256 below and the release manifest. The Windows executables are not Authenticode-signed. A matching hash checks file integrity; it does not certify that software is harmless or remove SmartScreen warnings.
  2. Open Cockpit and scanUse « Analyser ce PC ». Review the detected hardware and runtime. The website and public MCP cannot scan your computer.
  3. Test a model you already haveIn « Tests », check the exact installed model and the operation before confirming. If no model is installed, review any proposed download separately. Do not bypass your system's protections.
  4. Read your own measurementA failed or cancelled test is not a speed result. Keep your report locally; any sharing is a separate choice.

Follow a fixed benchmark recipe or troubleshoot local AI (French).

Download 0.1.1-beta

File names, build and hashes below come from the same release manifest as the French download page.

Windows .exe · 0.1.1-beta · build 20260928000500Download
SHA-256 353f68799d40d8129c7b586acd58542a0f334d3a7992e33dbcbbfb3e8a2f8968
Windows portable · 0.1.1-beta · build 20260928000500Download
SHA-256 e55d00ea2c4ee965bddd4a6e1868d8c3d013526d80e4d2a5c4fd301c19fbfd2d
Linux AppImage · 0.1.1-beta · build 20260928000500Download
SHA-256 2a275488c9e7175512b55a45467f051877b0b854f8883c7a421a1a38be218e00
Linux .deb · 0.1.1-beta · build 20260928000500Download
SHA-256 8841d200354b8b61f761284bb6b3acd85963d631b38617e90806239d4a349fe2
Linux .rpm · 0.1.1-beta · build 20260928000500Download
SHA-256 a1d1d3ad2fc9294a075850d415f815b3f21a8fa9f149cd7b83723792e379808b

Release manifest (JSON) · File verifier (French) · Known issues (French)

What this beta does not prove

The public evidence highlighted by OutilsIA comes from one physical RTX 4080 SUPER machine. It is not validation across every GPU, laptop or shared-memory PC. Linux packages are available; a complete external Linux user journey remains to be qualified.

Hermes 3 is a model. Hermes Agent is a separate runtime. Listing a model in the catalogue does not mean it can run locally: DeepSeek V4.1 Flash is a non-actionable watchlist entry, not a local installation promise.

The core scan, first benchmark and simple report are free. Founder checkout is closed. This page does not sell a licence or a subscription.

Help test another PC

We need independent Windows or Linux x64 tests, including PCs with 64, 96 or 128 GB of shared CPU/GPU memory. Start with a modest model, not an unverified huge download. The current kit does not cover Apple Silicon.

Prepare a private, redacted test report. Participation and publication are separate choices. No external result is counted before it has actually been run and reviewed.