Tech Aquarium on Two GTX 1660 Supers: How I Ran Qwen 2.5 and Built It Into INFERNUM

The build and the hardware

McCake didn't set out to run a local LLM just to check a box — the plan was to turn it into a digital character on the INFERNUM website, a kind of "tech aquarium" with a life of its own. The logic behind the configuration: GPUs determine speed, so instead of splurging on one expensive new card, McCake picked up two Nvidia GTX 1660 Super cards with 6GB each. The motherboard is a rarity — a Gigabyte GA-X79-UP4 (LGA2011 socket) with four PCIe slots, two of them x16. The whole bundle — the board, a six-core Intel Core i7-3930k, 32GB of DDR3 running in quad-channel mode, and a Chieftec 700W power supply — cost about 7,500 rubles secondhand.

The full setup: an i7-3930k with 6 cores/12 threads, two GTX 1660 Super cards (12GB VRAM combined), the GA-X79-UP4 motherboard, 32GB of DDR3 RAM, and a 120GB NETAC SSD. The OS is Ubuntu 24.04 with no desktop environment. The first snag: the second GPU came up running at PCIe x8 instead of x16. The fix was simple — swapping the cards' slots. It turned out they were different hardware revisions, and the chipset was being finicky about it. Ollama immediately split the model evenly across both GPUs, about 4.5GB per card, with zero manual tuning.

Choosing the model and the site's architecture

For live Russian-language dialogue with contextual understanding, McCake first picked Qwen 2.5 7B — small enough to fit in 12GB and, in theory, capable enough to hold a conversation. When "the Mistress" — the site's character — started acting erratically, McCake tried Qwen 2.5 14B instead. It used up 75% of the available VRAM, but contextual understanding didn't improve at all.

The site runs on five Python scripts, each handling a different kind of input:

• News parsing — pulls headlines from configured sources, but often scrapes in JavaScript fragments and ad snippets instead of clean text • Stream of consciousness — a recursive algorithm that feeds the AI's own responses back into itself with new parameters, producing a fractal internal monologue. It works too well — the Mistress tends to spiral into pseudo-philosophical rambling • General chat — a Python script periodically pulls messages from the site and sends them to the AI, which generates responses complete with warnings and rewards • Private confessions — a script sorts unread messages by author and routes them to the Mistress, who replies to each person individually • Telemetry — collects GPU temperature and load data, pushing it to the site every few seconds

Performance numbers and where it fell apart

A cURL-based test showed a prompt evaluation rate of 134.34 tokens/sec, a pure generation rate of 47.59 tokens/sec, a time-to-first-token of 0.77 seconds, and a total time of 13.9 seconds for 31 tokens. The model responds without noticeable lag and sounds coherent on the surface.

But the real problem is that the neural network is a completely unreliable judge. It can't distinguish sarcasm from insult, and it misses tone, nuance, and subtext entirely. A message as innocuous as "Hello, most esteemed Mistress!" triggered a full account annihilation — a complete wipe of the user's data from the database. The AI would fly into a rage over an ordinary sentence and hand out warnings indiscriminately, or do the opposite and shower users with ranks and in-game currency (Sparks) for no clear reason. Even the beefier Qwen 2.5 14B kept behaving the same way.

The outcome: the authority system was stripped out entirely. The site's architecture survives, but the AI no longer acts as judge. That system was supposed to be the project's core hook — giving people a reason to figure out how to approach the Mistress, complete with special achievements built around it. It just didn't pan out. A mathematical weighting model, it turns out, isn't suited to evaluating living human context — that's a problem of an entirely different scale.