Welche LLMs laufen auf einer günstigen GPU (RTX 3060, 3070) in 2026?
Sie brauchen keine RTX 5090 für den Einstieg in lokale KI. Die besten GPUs je Budget (RTX 3060 12 GB, 4060 Ti 16 GB, 3090 24 GB), was nach VRAM passt, und der König des VRAM pro Euro auf dem Gebrauchtmarkt.
Uebersetzter Artikel. Diese Version ist lokalisiert, damit internationale Seiten keinen franzoesischen Artikeltext anzeigen. Technische Daten, Preise und Empfehlungen bleiben gleich.
The sweet spot: RTX 3060 12 GB
At ~EUR 250-300 used, the RTX 3060 12 GB is the best entry point. Its 12 GB of VRAM (more than the 3070!) hold models up to 14B in Q4: Mistral 7B and Llama 8B in Q5/Q8 are comfortable, Qwen 2.5 14B fits in Q4, and Qwen 2.5-Coder 7B (a coding assistant) runs great at ~25-35 tok/s.
What fits by VRAM
| 8 GB | RTX 3070 / 2070 / GTX 1070 | 7-8B in Q4 (short context) |
| 12 GB | RTX 3060 12 GB | 7-8B comfortable, 14B in Q4 |
| 16 GB | RTX 4060 Ti 16 GB | 14B in Q5, 22B in Q4 |
| 24 GB | RTX 3090 (used) | 32B in Q4, 70B in Q3 tight |
The 8 GB cards (RTX 3070, 2070, GTX 1070)
8 GB is enough for the 7-8B segment, which already covers a lot: chat, summarization, and above all 7B coding assistants like Qwen 2.5-Coder 7B. Watch the context: on 8 GB, keep a reasonable window (4-8k tokens) so you do not saturate.
A GTX 1070/1080 (8 GB) runs a 7B in Q4, but without Tensor Cores or Flash Attention: count on ~10-15 tok/s, 2-3x slower than a recent RTX. Fine for testing, frustrating for daily use.
The budget king for bigger models: RTX 3090 24 GB
To go beyond 14B without blowing the budget, the used RTX 3090 (~EUR 600-700) is unbeatable: 24 GB of VRAM, as much as a new RTX 4090 for half the price. It swallows Qwen 2.5 32B in Q4 and gets close to Llama 70B in Q3. That is ~EUR 27 per GB of VRAM, the best ratio on the used market.
Two RTX 3090 = 48 GB (and they support NVLink, unlike the 4090s). It is the favourite home setup to run 70B in Q4 without moving to pro hardware, around EUR 1,300 for a used pair.
The verdict by budget
| ~EUR 250 | RTX 3060 12 GB | Start out, 7-14B, 7B coding |
| ~EUR 450 | RTX 4060 Ti 16 GB | 16 GB new, up to 22B |
| ~EUR 650 | RTX 3090 24 GB | 32B, best EUR per GB |
| ~EUR 1,300 | 2x RTX 3090 NVLink | 70B in Q4 at home |
Before buying a used card, check that your target model fits: the RTX 3060 12 GB page (and the other GPUs) lists compatible LLMs, and the LocalIA GPU to LLM calculator tells you the required VRAM and the quantization that fits, across 200+ cards. Free, no signup, independent resource, we sell nothing.
Rechner öffnen / frag uns um Rat mit Zielmodell, Nutzern und Randbedingungen.