The tenth best open source model in the world – and it runs on my desktop PC. A few weeks ago I combined two used graphics cards in one machine: twice twelve gigabytes of graphics memory, five year old hardware – and an AI model that beats models I would have to pay real money for.
The graphics card for that costs about 350 euros today. A year ago it was a hundred euros cheaper. The reason is simple: hardware is no longer built for people, it is built for AI.
The plan B before someone else turns off the light
Presidents ban access to models, manufacturers dictate their terms and peak usage hours are charged twice and three times over. The question is not whether local AI is perfect. The question is: what is your plan when someone else turns off the light?
Two graphics cards instead of a data center
Data center quality on old hardware
What Qwen 3.8 can really do on a five year old PC: surprisingly much. The model delivers results you would have to pay real money for at cloud providers.
Two cards in one machine
Running two graphics cards in one machine is easier than you think – and doubles the usable graphics memory. Hardware many of us already have at home.
Open source changes the equation
Open models cost nothing per request, run privately on your own hardware and do not depend on quotas, usage bans or price screws.
Why hardware prices keep rising
The big manufacturers no longer build their hardware for people, but for data centers. A current consumer graphics card easily costs several thousand euros. If you buy a used card today, you pay a hundred euros more than a year ago – not because it got better, but because AI demand has turned the market upside down.
The manufacturers' cycle
The pattern is always the same: test for free, use it cheaply, feel the limits. First the providers lure you in with free contingents, then come the limits, usage bans and price tiers for peak hours. That is exactly where local AI pays off: if you already have the hardware, you become independent of this cycle.
How useful local models have become on old cards is described in my article "Local AI is finally becoming useful".
Are you prepared?
The video shows compactly what is possible – from two graphics cards in one machine to the question that remains at the end: your own hardware or keep paying?
➡️ All content in the embedded video was created with AI.