One question: what size of local AI actually fits this PC for real jobs. Not "which model is best in the world." Pictures below are the real screens. Click any picture to enlarge it.
1
Start
Open the app and land on Home
First you see a splash while Doctor checks this PC. Wait. You do not need a terminal. Then Home is the map: Home, Doctor, Models, Benchmark, Reports, Compare, Settings. The life-buoy opens this help.
Opening screen
Splash. Wait here while Doctor checks this PC. You do not need a terminal.
Home
Home. Orange button is Guide me through. Bottom row is your journey: Hardware → Models → Roles → Run → Report.
When Home looks ready, click Guide me through. The rest of this help follows that path.
2
Doctor
Let Doctor check this PC
Doctor confirms Windows pieces, the local model runner (Ollama), storage, and that acceleration is real.
While it runs you may see "Checking this PC…" — wait; do not close the app.
Needs attention — something failed. Use the repair button on that page. Stay in the app.
Everything is ready — you can benchmark. The header pill should say Ready before a long run.
Ready
Doctor — ready. Green checks. You can re-run checks from this page anytime.
3
Models
Download or pick a model
Open Models. Packs live on this PC through Ollama. Scoring does not send answers to a public chat website. One model per report.
Installed — empty
Installed, empty. Switch to Preferred or Search Ollama and download one.
Preferred
Preferred. A short roster we suggest for this PC. Download from here if you are not sure what to pick.
Search Ollama
Search Ollama. Type a name if you already know which pack you want.
Downloading
Downloading. Leave this window open until it finishes. Then it appears under Installed.
Installed — with models
Installed. Tick the one this run should test.
4
Guide me through
The guided setup (intro, then work)
From Home, Guide me through fills this window. Most steps come in a pair: an illustrated intro with OK, let's go, then the working screen. If nothing is installed, you only get the first picture below.
If no model is installed
Need a model first. Search or Go to Models, download one, then start the guide again.
Step 1 intro — OK, let's go
Intro. "Choose the model to test." Click OK, let's go.
Step 1 work — pick the model
Work. "Which model should we test?" Pick one installed row, then Next.
Step 2 intro — OK, let's go
Intro. "Choose your personas." A persona is a real customer role.
Step 2 work — roles
Work. Home / Play / Work. Keep all 20, or clear a group and tick only what you care about. Each role has its own idea of "fast enough."
Step 3 intro — OK, let's go
Intro. "Thinking on or off?" Off is the usual default.
Step 3 work — thinking
Work. Thinking Off answers sooner. Thinking On can take much longer — the bench waits until thinking is done.
Thinking Off
First pass on everyday speed.
Thinking On
Only when careful reasoning matters more than a snappy reply.
Step 4 intro — OK, let's go
Intro. "Save the final report (optional)." The app already keeps a copy.
Step 4 work — save and start
Work. Optional extra folder, then Start benchmark.
5
Benchmark
Watch the run — or pause / stop
You can also start from the Benchmark tab without the guide. Progress, RAM / VRAM / CPU, and the current question show live. Stop still keeps whatever finished.
Idle — not started yet
Ready to benchmark. Start Benchmark if you skipped the guide.
Running
In progress. Let the PC work. A full run can take a while on a large model.
Paused
Paused. Resume when you are ready, or Stop to keep a partial report.
6
Reports
Open what this PC produced
If you have not finished a run yet, Reports says there are no runs stored — that is normal. After a run, read the match labels first (Excellent, Strong, and so on). Ask: did this size feel usable for the roles I picked, on this hardware?
Library
Your benchmark library. Open a row, Compare, or Delete.
One report
Opened from the list. Open report, PDF, Folder, or Delete this run.
Choose what to remove
Delete picker. Tick the runs to remove from this PC, then Delete.
Confirm
Are you sure? This completely removes those reports from this PC.
7
Compare
Put two (or more) reports on one page
Compare does not run the bench again. It builds one page from reports you already have — for example a small model vs a larger one on this PC. If the list is empty, finish at least two benchmarks first (or upload saved HTML).
First visit — OK, let's do it
Intro. Same PC. Different LLM sizes. Click OK, let's do it.
Pick reports
Work. Tick reports from at least two different models, then Run comparison.
8
Settings
Folder, color, updates
Color theme changes the whole app, including this help page. Layout stays the same; only the palette changes.
Preferences. Extra report folder, Haval / Classic / Dawn / Flexoki, and Update.
How judgment and scores are decided
Scoring is a funnel. We do not average a slow disaster with a pretty paragraph and call it a win.
Did it finish? If it never produced an answer, that is Failed — not "slow."
Was it far too slow for that role? If yes, it is Not recommended even if the text looks excellent.
Combined score. About half is everyday role work (Phase 1). About half is a skill pack (Phase 2). Skills are weighted by the role — an engineer needs coding; a family coordinator does not.
Role floors. Some jobs refuse a label if a required skill is too weak.
Final match label from the scale below.
The funnel
Finish first, then speed, then the 50 / 50 mix, then role floors, then a label. Click to enlarge.
Excellent 88 and up
Strong 75–87
Acceptable 62–74
Marginal match 48–61
Not recommended below 48, or blocked by speed / a role floor
Failed no usable finish
Size is not the winner. Fit is.
How the app is put together (short)
You see one window. On this PC, helpers work together locally: