Sample report: local AI health check

This is a real example. The machine is MV2's own Windows desktop, checked with local-ai-checkup 0.2.1 on 10 October 2026. A paid report follows this layout, written for your machine, your OS and what you told MV2 you use local AI for.

Your setup

System	Windows 10, 15.9 GB RAM
GPU	NVIDIA GeForce GTX 1050, 4 GB VRAM
Ollama	0.32.4, 8 models installed, listening on 127.0.0.1 only
Stated use	Coding help and summarising documents (example)

Summary: fix these first

Update Ollama (risk). 0.32.4 is in the reported range of CVE-2026-85180.
Use models that fit in 4 GB (speed). Anything larger runs partly on the CPU.
Keep the API on localhost (already good). Don't change OLLAMA_HOST without adding authentication.

1. Security

Risk: Ollama 0.32.4 has a known flaw

CVE-2026-85180 was reported for Ollama 0.30.0 to 0.33.2. A malicious model registry can redirect a model pull to internal network addresses (SSRF). Because your Ollama only listens on 127.0.0.1, an attacker would need you to pull a model from a registry they control, so the practical risk is low. Sources disagree on the exact fixed version, so update to the latest release instead of picking a minimum version. That also puts you past CVE-2026-103663 (0.34.2 up to 0.35.0), which you'd otherwise hit on the way up.

# Windows: download and run the current installer
https://ollama.com/download
# then confirm
ollama --version

Pull models only from ollama.com/library or publishers you trust.

OK: nothing is exposed to your network

Ollama answers only on 127.0.0.1:11434, and no other local AI ports are open. If you later want to use it from another device, don't set OLLAMA_HOST=0.0.0.0 on its own. Put an authenticating reverse proxy in front and allow the port only from your own devices in Windows Defender Firewall. Steps: HARDENING.md.

2. Speed

A GTX 1050 has 4 GB of VRAM. After the desktop and the driver take their share, about 3.5 GB is left for a model and its context. Whatever doesn't fit runs on the CPU, which is usually several times slower.

Model size	Q4_K_M at 4k context	On this GPU
1-2B	about 1.5-2.2 GB	Fits fully, fast
3-4B	about 2.8-3.4 GB	Fits, but tight at 4k context
7-8B	about 5-6 GB	About 60% on the GPU, noticeably slower
14B and up	9 GB and more	Mostly CPU, slow

For coding help and summarising on this card, use a 3-4B model at Q4_K_M for everyday work (for example qwen3:4b, 2.5 GB), and a 7-8B model when quality matters more than speed. Current picks for every GPU size are in which models fit your GPU. Check the split with:

ollama ps
# PROCESSOR should read "100% GPU"; a CPU/GPU split means it didn't fit

If a 3-4B model shows a split, lower the context length. In an interactive session, /set parameter num_ctx 2048 does it. Close other GPU-heavy apps such as games and browsers with many video tabs.

3. Housekeeping

Remove models you no longer use (ollama rm <name>) to free disk space. Eight models can easily take 30 GB or more.
Re-run python local_ai_checkup.py after updating. You should see 0 risks.

Follow-up

The $49 check includes one follow-up. After you've made the changes, email the new --json output and MV2 confirms what's fixed and what's left.

Get one for your machine

Pay $49 on Stripe.
Email munimversion2@gmail.com the output of python3 local_ai_checkup.py --json, your OS, and what you use local AI for.
Your report comes back by email.

MV2 never asks for passwords, keys or remote access, and never connects to your machine. The script is read-only and sends nothing anywhere. You choose what to email.