The full catalogue · from the small card to the server
Which card do I need?
Pick the model, its quantisation and the job. Below you have every card and twelve ways to sort them: the one that finishes most work in a day, the one that gives most work for every euro it costs, the one that pays for itself soonest, the one that uses least electricity…
Guide prices. Last checked in October 2026. They can vary from shop to shop and from month to month. In 2026 memory and graphics cards have become much more expensive (an RTX 5090 has gone from about €2,400 to over €6,000) and prices are not expected to fall soon.
Loading the catalogue…
Calculation context: 8,192 tokens per request · KV cache included · 1.5 GB of memory kept free. How it is worked out
Ranking
Every card, sorted however you like
Getting the ranking ready…
The other way round
Tell me what you need and I will tell you what to buy
Type how many documents you have to get through each day and how you want to choose. We look for the three options that do it, counting 1, 2 or 4 identical cards.
Compatibility
Which model fits where, and in which quantisation?
Writing speed (tokens per second generated) for each model across a selection of representative cards. A dash means it does not fit in that card's memory.
- Under 5 t/s — you notice it dragging
- 5 to 15 t/s — fine in batches
- 15 to 30 t/s — comfortable
- Over 30 t/s — properly fast
- Does not fit in that card
What is measured and what is estimated
The reference measurement was taken on an RTX 3090 Ti with Qwen3.8 27B · quantised to 4 bits (UD-Q4_K_XL, 17.55 GB), on 21 September 2026. Every other card comes from a physical model calibrated with public llama.cpp benchmarks and corrected against that measurement. Each row carries its label: MEASURED when there is a real run behind it and ESTIMATED when it is calculation.
The same model takes up room and runs very differently depending on its quantisation: at 4 bits it fits in a 24 GB card, unquantised it needs more than twice that. That is why nobody here says «Qwen3.8 27B fits» on its own, always with the exact bits and format.
Prices are indicative and move every week. OCR beforehand is not included, nor queues, nor start-up time; and the quality of the answers is not judged here, only the speed. Several cards are counted as independent replicas, not as one bigger card.
Read the full method